# Arinti > Arinti is an AI-native consultancy delivering production-grade AI and data solutions for enterprise and public sector organisations across Belgium. Arinti NV is based in Kontich, Belgium (Veldkant 35C, 2550). Contact: hello@arinti.ai or via https://arinti.ai/contact. This file contains the full text content of https://arinti.ai for LLM consumption. A compact index (including Dutch URLs) lives at https://arinti.ai/llms.txt. ## Pages ### [About Us](https://arinti.ai/about-us) WE WORK WITH YOU NOT FOR YOU HOW WE OPERATE #### Doing what drives impact. We start by understanding your business and defining a strategy together. We don't do technology for technology's sake. By understanding your goals and identifying where AI can truly make a difference, we unlock business value that matters. Focusing on the drivers that actually impact your organisation, we help you strengthen and consolidate your competitive position. OUR TRACK-RECORD #### Proven track record and cross-sector experience that delivers. With more than seven years of experience, we bring expertise that predates the current wave surrounding AI. We build real-world AI solutions that don't get stuck in the prototype phase. We deliver production-ready systems that we manage and scale together with our clients. From helping Unilever forecast sales to enabling Stad Kortrijk to launch their GenAI virtual assistant for their citizens. CAPABILITIES #### Clear data first, smart processes next. Our strength lies in our combined expertise in data and AI engineering. Successful AI projects don't start with models, but with a pragmatic data strategy and structured access to high-quality data. We build a solid, reliable data foundation first, then develop innovative AI applications that help you achieve your goals. AI CODING --- ### [Careers](https://arinti.ai/careers) WANT TO JOIN OUR TEAM? GROW WITH US #### Join a team at the frontier of data and AI We're always looking for people who want to do meaningful work in AI for clients that matter. Not chase trends, but build production-grade solutions that move organisations forward. You'll work alongside a team that's curious, sharp, and collaborative. What matters most is the right attitude: intellectual curiosity, a bias for action, and the drive to keep learning in a field that never stands still. CULTURE #### How we work at Arinti Experience a human-sized company, small enough that your work matters, structured enough that it ships. Our culture is built on transparency and pragmatism, towards our clients and towards each other. No politics, no layers of abstraction, just honest conversations about what works and what doesn't. We're a group of like-minded people devoted to technology, working at the frontier of data and AI. We look beyond the hype to deliver pragmatic solutions built on what really works. Simple, honest, very Arinti. --- ### [Cases](https://arinti.ai/cases) WHAT REAL IMPACT LOOKS LIKE clients #### Built on strong partnerships --- ### [Cookie Policy](https://arinti.ai/cookie-policy) Cookie Policy #### Cookie Policy #### Who are we In this Cookie Policy, "we", "us" or "our" means Arinti NV, with its registered office at Veldkant 33A, B-2550 Kontich and with company number BE 0671.586.428. We act as controller for the personal data we gather through your use of our website [https://arinti.ai/](https://arinti.ai/). Since our Website makes use of "cookies", we would like to inform you about the different types of cookies we use and which cookies our service providers use and how you can manage and block cookies. If you have any questions, concerns or complaints regarding this Cookie Policy or our use of cookies or you would like to revoke your consent, you can contact our privacy department: Via-mail: [hello@arinti.ai](mailto:hello@arinti.ai) to the attention of our Data Protection Officer By post to: Arinti NV, To the attention of the DPO, Veldkant 33A, 2550 Kontich – Belgium This Privacy Policy was revised last on 04/12/2023 #### What are cookies? Cookies are small pieces of data that a website asks your browser to store on your computer, laptop or mobile device when you visit a website or certain pages of that specific website. The cookie allows that website to "remember" your actions or preferences over time and to apply them during subsequent visits, for example your preferred language. Cookies usually contain the name of the website where the cookie has come from, how long the cookie will remain on your device (its expiry date), and a value which is usually a randomly generated unique number. A distinction is made between strictly necessary cookies and other cookies. The strictly necessary cookies are necessary for the proper functioning of the Website, for example for the security of our Website and its accessibility. By virtue of the applicable legislation, we are not required to ask your consent to use such cookies. A further distinction on the basis of the category can be made according to the specific purpose of the cookie. For each of these categories (unless they are strictly necessary), we will ask your consent. Cookies can be divided into the following categories: Permanent cookies and session cookies: The task of a permanent cookie is to recognize the web browser with a view to repeat visits to the Website, while the task of a session cookie is to recognize the web browser during the visit to the Website. Permanent cookies will remain stored on your computer or mobile device. Session cookies will be deleted at the end of every session. First party cookies and third-party cookies: First party cookies are set by the Website itself. Third party cookies are set by other websites than our Website. This happens when our Website incorporates elements from other websites, such as social media plug-ins or advertisements that are put on our Website. Performance cookies and functional cookies: Performance cookies monitor how the visitor uses the Website. This enables the website administrator to adjust and optimize the content and ease of use of the Website (for example: how many people have visited the different pages on the Website and which language is the most popular amongst our visitors). Functional cookies are cookies that remember your preferences when you use the Website, such as your chosen language, and will make the Website more user-friendly and enhancing your user experience. Cookies will not always contain personal data but may contain information that may lead to your identification, in which case such cookie shall be treated as personal data. To the extent a cookie can be considered as personal data, our Privacy Policy will apply. #### Which cookies do we use? CONSENT | Analytics | YouTube sets this cookie via embedded youtube-videos and registers anonymous statistical data. | 16 year 2 months prism_67322792 | Other | prism.app-us1.com | 1 month prism_67322792 | Other | .arinti.ai | 1 month YSC | Advertisement | YSC cookie is set by Youtube and is used to track the views of embedded videos on Youtube pages. | session VISITOR_INFO1_LIVE | Advertisement | A cookie set by YouTube to measure bandwidth that determines whether the user gets the new or old player interface. | 5 month and 27 days test_cookie | Advertisement | The test_cookie is set by doubleclick.net and is used to determine if the user's browser supports cookies. | 15 minutes IDE | Advertisement | Google DoubleClick IDE cookies are used to store information about how the user uses the website to present them with relevant ads. | 1 year and 24 days yt-remote-device-id | Advertisement | YouTube sets this cookie to store the video preferences of the user using embedded YouTube video. | Never yt-remote-connected-devices | Advertisement | YouTube sets this cookie to store the video preferences of the user using embedded YouTube video. | Never #### Cookies used by our partners Our Website makes use of social media plugins. These social media partners use third party cookies. The following social media partners are used: Twitter and Facebook. Links to relevant cookie policies: - Twitter: [https://help.twitter.com/en/rules-and-policies/twitter-cookies](https://help.twitter.com/en/rules-and-policies/twitter-cookies) - Facebook: [https://www.facebook.com/policies/cookies/](https://www.facebook.com/policies/cookies/) - Google+: [http://www.google.com/policies/technologies/cookies/](http://www.google.com/policies/technologies/cookies/) - YouTube: [https://policies.google.com/technologies/cookies?hl=en](https://policies.google.com/technologies/cookies?hl=en) #### Managing and blocking cookies When visiting the Website visitors will be informed (by means of a pop-up) which cookies are being used by us, and we will ask the visitor's consent for the use of each of the cookies. You can at all times manage or delete cookies either directly on this website or by using the settings on your internet browser, allowing you to refuse some or all cookies, except for the strictly necessary cookies. Turning off cookies will limit the service that we are able to provide and may affect your user-experience. Deleting cookies may result in manually adjusting preferences every time you visit our Website. For more details about controlling and/or deleting cookies, visit: - Chrome: [https://support.google.com/chrome/answer/95647?hl=en](https://support.google.com/chrome/answer/95647?hl=en) - Safari: [https://support.apple.com/kb/PH21411](https://support.apple.com/kb/PH21411) - Internet Explorer: [https://support.microsoft.com/en-gb/help/17442/windows-internet-explorer-delete-manage-cookies](https://support.microsoft.com/en-gb/help/17442/windows-internet-explorer-delete-manage-cookies) - Firefox: [https://support.mozilla.org/en-US/kb/enable-and-disable-cookies-Website-preferences](https://support.mozilla.org/en-US/kb/enable-and-disable-cookies-Website-preferences) - Opera: [http://www.opera.com/help/tutorials/security/cookies/](http://www.opera.com/help/tutorials/security/cookies/) - Edge: [https://privacy.microsoft.com/en-us/windows-10-microsoft-edge-and-privacy](https://privacy.microsoft.com/en-us/windows-10-microsoft-edge-and-privacy) #### Amendments to the Cookie Policy We reserve the right to make amendments to this Cookie Policy. The most recent version of the Cookie Policy can be consulted on our Website. --- ### [Home](https://arinti.ai/) DATA ENGINEERING AI STRATEGY Chatbots GEN AI AI ENGINEERING AGENTIC PROCESS AUTOMATION AGENTIC AI DATA ENGINEERING AI ROADMAP MACHINE LEARNING AI ADVISORY AI MATURITY ASSESMENT DATA GOVERNANCE DATA ARCHITECTURE FROM AI AMBITION TO REAL-WORLD IMPACT Arinti is an AI-native organisation covering the full spectrum from strategy and data to innovative GenAI applications, with the expertise and discipline to make it work. Our experience goes back well before the recent surge in GenAI. We don't advise from the sidelines, we're in the trenches of your most challenging AI projects and stay involved until there's measurable business impact. Watch our story Vision #### On the frontier of AI, focused on measurable outcomes Arinti is an AI-native organisation covering the full spectrum from strategy and data to innovative GenAI applications, with the expertise and discipline to make it work. #### From AI ambition to real-world impact Cases #### Insight-driven work built with purpose DATA & AI STRATEGY EXERCISE #### Is your organisation ready for data & AI? Six pragmatic deep dives that turn departmental needs into one executable roadmap. Start with a free 2-minute readiness check. READINESS CHECK Clients #### Built on strong partnerships Services #### How we can help your organisation APPLIED AI REAL IMPACT Testimonials #### Thoughts from our clients Insights #### Curious about our opinion? --- ### [Insights](https://arinti.ai/insights) INSIGHT-LED WORK --- ### [Privacy Policy](https://arinti.ai/privacy-policy) Privacy Policy #### Privacy Policy Arinti cares about your privacy and always acts in this respect in accordance with the stipulations of privacy legislation, including the General Data Protection Regulation (GDPR). This privacy statement wishes to inform you about the processing of your personal data when visiting our website [https://arinti.ai/](https://arinti.ai/) (hereinafter referred to as "the Website"), making use of our services, disclosure and transfer of your personal data, and the rights you have with regards to the processing of your personal data. For our privacy practices in relation to our services, we would like to refer to the agreement as may be conducted between us in addition to this privacy statement. #### Who are we? "We", "us" or "our" means: Arinti NV, with its registered offices at Veldkant 33A, B-2550 Kontich and with company number BE 0671 586 428. We act as controller for the personal data we gather through your use of our website. If you have any questions, concerns or complaints regarding this Privacy Statement or our processing of your personal data or you wish to submit a request to exercise your rights, you can contact us: Via-mail: [hello@arinti.ai](mailto:hello@arinti.ai), to the attention of our Data Protection Officer By post to: Arinti NV, To the attention of the DPO, Veldkant 33A, 2550 Kontich – Belgium #### Why do we process your personal data? We collect and process your personal data for the following purposes: - Individuals who visit our website - To interact with you when you contact us - To send you our newsletter - For direct marketing purposes - For client administration purposes (CRM) - Individuals using our application - Individuals whose personal data we obtain in connection with providing services to our clients - To fulfil our obligation set out by law #### What personal data do we process? We primarily process your personal data when you give it to us (e.g. when visiting our website or contacting us). In addition, it is possible that we process your data because other people give it to us (e.g. your employer while professional our professional services). Where we need to process personal data to provide professional services, we ask our clients to only provide the necessary information to the data subjects regarding its use. When you visit our website, we process the personal data that you provide voluntarily, e.g. when completing online forms or subscribing to a newsletter. The following data can be processed: - Identification data (e.g. first and last name) - Contact data (e.g. e-mail address and phone number) - Professional data (e.g. job title, education, company) - Information required for fulfilling our services to you - Any other personal data that you voluntarily choose to share with us We will never process any sensitive categories of personal data unless you provide us with such data. If there are any free text boxes you are not required, and should not disclose, any type of sensitive personal information. We also process certain personal data automatically when you visit our website. This data includes your IP address, pixel ID, device type, broad geographic location, etc. In addition, our website also features certain social media plugins for Facebook and LinkedIn. When you commission us to provide you with professional services, we process personal data when we have a compelling reason to do so with regards to those services. In the context of providing professional services to customers, we also process personal data of individuals that are not directly our clients (e.g. employees, suppliers, etc.). The personal data that we process to provide our services is offered voluntarily by (or collected by us from third-party sources at the request of) our clients. This information can include: - Identification data (e.g. first and last name) - Contact data (e.g. e-mail address and phone number) - Professional data (e.g. job title, education, company) - Financial information (e.g. payment information) - Personal data relating to you or third parties that you provide us for the purpose of fulfilling our services This personal data is used to provide you with the requested service, to administer and maintain contractual relations, marketing and business development, comply with legal obligations, exercise or defend legal rights, and/or accounting and tax purposes. When you want to make use of our contact form we process the following personal data: - Identification data: first and last name - Contact data: e-mail address and phone number - Professional data: function and company We also process any personal data that you choose to put in the designated blank field (please do not provide us with any sensitive information, such as health information, information pertaining to criminal convictions, or credit card/account numbers). If you subscribe to our newsletter you will also be asked to provide your e-mail address. #### Legal ground for processing your personal data We process the personal data we receive from you via our contact form or subscription to our newsletter based on "consent". We use the personal data we receive from you via our contact form based on your implied consent. The user information we aggregate from your behavior on our website is processed in accordance with our legitimate interest. Your user information is used to make our users' website experience as comfortable and safe as possible. In any case, we strive towards a balance between the legitimate interest and respecting your privacy. Your personal data will also be used for direct marketing if you gave an explicit consent to do so. If you are already added to our mailing list to receive electronic marketing material, we can use your data to send marketing regarding our services. In the event the legal basis for the processing of your personal data is consent, you will, at all times, have the right to withdraw your consent. This will, however, not affect the lawfulness of any processing done prior to the withdrawal of consent. And in the event the legal basis for processing is our legitimate interest, you will have the right to object to such processing, as set out below. #### Disclosure of your personal data to others We only share your personal data with third parties if necessary, to carry out one or more of the as forementioned processing activities. To maintain and improve our services we sometimes invoke third parties, e.g. developers and service providers. Where necessary for the services provided by these third parties, your personal data may be transferred to them. We conclude contractual agreements with these third parties to ensure that they only process your personal data in accordance to our instructions and that they always meet the same degree of security and confidentiality as we do. We use the following categories of service providers: CMS System, Google Analytics. We may also disclose your personal data in the event such disclosure is required or necessary to fulfil a legal obligation. This Privacy statement is only applicable to us and our Website. The Website can include links to other websites, which we do not monitor. Subsequently, this Privacy Statement is not applicable to those websites. We will not transfer your personal data outside of the European Economic Area. #### Retention of your personal data We commit ourselves not to retain your personal data longer than necessary for the specific processing activity or any retention period that is required by law. #### Protection of your personal data We make every effort to (continue to) guarantee the accessibility, confidentiality, and integrity of your personal data. To this end, we have taken the necessary steps to protect your personal data. This means, among other things, that the number of employees who have access to your personal data is carefully selected and strictly limited. In addition, the impact on your privacy is always analysed in detail to guarantee your rights, the security and protection of your personal data. In the event of an unforeseen incident involving your personal data, Arinti will always inform you in accordance with the legal provisions. #### Your rights and how to exercise them To exercise any of your rights, please send us a written request through one of the channels mentioned at the top of this Privacy Statement. We will respond to your request without undue delay, but in any event within one month of the receipt of the request. The right to access: You have the right to access all personal data that we process about you. You can also receive a copy of this data, which we will send to you on paper or digitally. You can do this by contacting our DPO by sending an email to [hello@arinti.ai](mailto:hello@arinti.ai). The right to rectification: If the personal data we hold about you is inaccurate or incomplete, you have the right to have this information rectified or, considering the purposes of the processing, completed. The right to erasure (right to be forgotten): You have the right to request the removal of your personal data if they are no longer necessary in the light of the purposes of this privacy statement or if you withdraw your consent to the processing. You can do this by contacting our DPO by sending an email to [hello@arinti.ai](mailto:hello@arinti.ai). Please note: We are not obliged to remove your personal data if we have to or may keep that data because of a legal or contractual obligation. The right to restrict processing: You have the right to restrict the processing of your personal data (meaning that the personal data may only be stored by us and may only be used for limited purposes), if: - You contest the accuracy of the personal data (and only for as long as it takes to verify that accuracy) - The processing is unlawful, and you request restriction (as opposed to exercising the right to erasure) - We no longer need the personal data for the purposes of our processing, but you require personal data for the establishment, exercise or defense of legal claims - You have objected to processing, pending the verification of that objection In addition to our right to store your personal data, we may still otherwise process it but only: with your consent; for the establishment, exercise or defense of legal claims; for the protection of the rights of another natural or legal person; or, for reasons of important public interest. We will inform you before we lift the restriction of processing. The right to data portability: To the extent that the legal basis for our processing of your personal data is consent, and such processing is carried out by automated means, you have the right to receive your personal data from us in a structured, commonly used and machine-readable format. However, this right does not apply where it would adversely affect the rights and freedoms of others. You also have the right to have your personal data transferred directly to another company, if this is technically possible, and/or to store your personal data for further personal use on a private device. The right to object to processing: If you feel that we are not using your personal data correctly, you may object to such processing, such as processing in the context of newsletters or other processing that is not necessary in the context of our products and services. You can do this by contacting our DPO by sending an e-mail to [hello@arinti.ai](mailto:hello@arinti.ai). The right to complain to a supervisory authority: If you consider that our processing of your personal information infringes data protection laws, you have a legal right to lodge a complaint with a supervisory authority responsible for data protection. You may do so in the EU member state of your habitual residence, your place of work or the place of the alleged infringement. In Belgium, you can submit a complaint to the Authority for the protection of personal data (Data Protection Authority), Drukpersstraat 35, 1000 Brussel ([contact@apd-gba.be](mailto:contact@apd-gba.be); [https://www.dataprotectionauthority.be/contact-us](https://www.dataprotectionauthority.be/contact-us)). #### Amendments to the Privacy Statement We reserve the right to modify this Privacy Statement at any time, and without prior notice, by posting an amended Privacy Statement on the website. You will always be able to consult the most recent version of the Privacy Statement on the website. This Privacy Statement was revised last on 03/08/2021, version nr. 2021/01. --- ### [Services](https://arinti.ai/services) THIS IS WHAT WE DO ADVISORY #### From AI ambition to real impact AI only delivers value when it's connected to what matters most. Together with your leadership team, we identify the challenges and opportunities where AI will have the greatest impact, then turn them into a prioritised roadmap backed by clear business cases. AI capabilities are mapped, and readiness is evaluated across people, processes, data, and infrastructure. Our structured evaluation framework ensures every initiative on the roadmap is both ambitious and achievable. #### AI BUSINESS ADVISORY A prioritised roadmap backed by clear business cases. Vision without the right foundation leads nowhere. We help you build an AI Center of Excellence that actually works by assessing your current AI maturity and capabilities, defining the right strategies, and guiding the technical choices that will make or break your AI ambitions. Together, we define your future state, chart the path to get there, and make sure today's decisions don't become tomorrow's bottlenecks. Pragmatic and tailored to your reality. #### AI TECHNICAL ADVISORY STRATEGY EXERCISE #### First get your data sharp, and your processes smart. A pragmatic discovery exercise: six deep dives, one executable Data & AI roadmap. ADVISORY DATA #### Data foundations that make AI possible Clear data first, smart processes next. Your data platform should be high-performing, efficient, and built to grow. We design, build, and maintain the infrastructure your organisation needs to turn data into a reliable, scalable asset. That means robust pipelines that automate ingestion, processing, and preparation. We bring diverse sources together into unified, well-structured datasets ready for analytics and AI. #### DATA ENGINEERING Unified, well-structured datasets ready for analytics The best models are the ones that actually get used. That's why we work in close collaboration with your business stakeholders, from initial exploration all the way to production-grade models integrated into your operations. Predictive modelling, classification, clustering, anomaly detection, recommendation engines, operational optimisation, applied to domains like demand forecasting, churn prediction, fraud detection, and predictive maintenance. Data becomes valuable when it drives better decisions and smarter actions. #### DATA SCIENCE & MACHINE LEARNING DATA ENGINEERING #### AI that delivers from day one, and grows from there The gap between a promising experiment and a production-grade solution is where most AI initiatives stall. We engineer AI solutions that are built for production from day one. Our motto: build something small, valuable, and measurable, then scale further from there. Our approach covers the full lifecycle: sourcing and preparing high-quality data, developing and fine-tuning bespoke models, building and integrating custom AI solutions, and deploying in the cloud, on-prem, or at the edge, fully integrated with your existing systems and IT landscape. #### AI ENGINEERING The gap between a promising experiment and a production-grade solution Agentic AI represents the next frontier: AI systems that don't just respond to queries but take initiative, reason through multi-step tasks, and act autonomously within your business processes. Think specialised, always-available teammates: resolving customer inquiries 24/7, surfacing real-time market insights, screening candidates at scale, or automating complex operational workflows. #### AGENTIC PROCESS AUTOMATION What makes this work reliably at scale is engineering discipline: robust governance, guardrails, and monitoring built in from the start. Whether you begin with a single agent automating one process or build toward a mesh of specialised agents across your organisation, the path from concept to production stays pragmatic and value-driven. ENGINEERING CLIENTS #### Built on strong partnerships TESTIMONIALS #### Thoughts from our clients --- ### [Terms of Use](https://arinti.ai/terms-of-use) Terms of Use #### 1. General These Terms of Use govern your access to and use of the website of Arinti NV, located at arinti.ai. By accessing and using this website, you accept and agree to be bound by these terms. #### 2. Intellectual property All content on this website, including but not limited to text, graphics, logos, images, software, and the design and arrangement thereof, is the property of Arinti NV and is protected by Belgian and international copyright, trademark, and other intellectual property laws. You may not reproduce, distribute, modify, or create derivative works from any content on this website without our prior written consent. #### 3. Limitation of liability The information on this website is provided "as is" without any warranties, express or implied. Arinti NV does not guarantee the accuracy, completeness, or timeliness of the information provided. To the fullest extent permitted by law, Arinti NV shall not be liable for any direct, indirect, incidental, or consequential damages arising from your use of or inability to use this website. #### 4. External links Our website may contain links to third-party websites. These links are provided for your convenience only. We have no control over the content of these sites and accept no responsibility for them or for any loss or damage that may arise from your use of them. #### 5. Applicable law These Terms of Use are governed by and construed in accordance with Belgian law. Any disputes relating to these terms shall be subject to the exclusive jurisdiction of the courts of Mechelen, Belgium. #### 6. Modifications We reserve the right to modify these Terms of Use at any time. Changes will be effective immediately upon posting on this page. Your continued use of the website after any changes constitutes your acceptance of the new terms. #### 7. Contact For questions about these Terms of Use, please contact us at [hello@arinti.ai](mailto:hello@arinti.ai). Last updated: March 2026 ## Current offerings ### [Data & AI Strategy](https://arinti.ai/data-ai-strategy) > A pragmatic Data & AI strategy exercise: six focused deep dives and an AI readiness assessment that turn departmental needs into one executable Data & AI roadmap in about twelve weeks. 6 FOCUSED PHASES ±12 WEEKS, START TO ROADMAP ½-day SESSIONS, ON-SITE 1 EXECUTABLE STRATEGY DATA & AI STRATEGY EXERCISE #### First get your data sharp, and your processes smart A pragmatic discovery exercise that turns departmental needs into one executable Data & AI roadmap. People and processes first, technology second. #### A pragmatic roadmap Structural foundations combined with feasible quick wins, not a slide deck that gathers dust. #### A scalable data foundation A foundation that grows with your organisation and stays simple to manage and integrate. #### Concrete recommendations A phased approach that couples fast impact to strategic build-up, prioritised on business value. #### Governance in the lifecycle Building on your existing AI policy, bringing continuity to every data and AI initiative. WHY THIS EXERCISE #### From scattered experiments to one shared direction Most organisations don't lack AI ambition, they lack alignment. Experiments happen in pockets, data lives in silos, and value stays anecdotal. Through a series of focused deep dives and workshops, we connect departmental needs with data and AI opportunities. As a vendor-neutral partner, we deliberately decouple exploration from execution: first insight, then decisions. Cross-departmental data opportunities Gap analysis: current vs. desired situation Quick wins and longer-term recommendations #### Deep dive · Business & departments Understand how data and AI are used today in each department, and identify where the biggest value sits, including wins possible without AI, through process optimisation. Governance & data-quality model Architecture scope focused on quick wins Data-quality roadmap and tooling advice #### Deep dive · Data & technical foundation Define a pragmatic data foundation that guarantees quality, availability and accessibility. First the data in order, then the tools. Prioritised AI opportunities (impact × feasibility) Quick wins and pilot candidates Governance anchored in the delivery lifecycle #### Deep dive · AI, governance & lifecycle Identify and prioritise AI use cases with measurable value; assess the AI-readiness of data and systems, building on your existing AI policy. Documented findings across all deep dives #### Validation session All insights from the deep dives, structured per domain and validated with the experts involved, so the joint workshop starts from a supported base. Unified strategic vision Implementation roadmap with pragmatic milestones Resource & capacity plan #### Joint workshop · Data & AI alignment Bring all insights together with management of business, IT and operations into one coherent, executable data and AI strategy. Full strategy document with roadmap and recommendations #### Final delivery One complete, pragmatic strategy document your organisation can use independently for decision-making and execution. THE APPROACH #### Six pragmatic deep dives, one coherent strategy Result-driven sessions that bring business, operations and IT together, doubling as a pragmatic AI maturity assessment. Every session is half a day on-site, with preparation upfront and synthesis afterwards, and every session delivers tangible output. * Indicative timing, aligned with stakeholder availability. 6 STATEMENTS 2 MINUTES NO EMAIL a Business & Departments deep dive, mapping where data and AI create real value in your core processes. the Data & Technical Foundation deep dive: first get the data in order, then the tools. the AI, Governance & Lifecycle deep dive, prioritising use cases on feasibility and business value. Departments know which data and AI use cases would create value in their core processes. Reporting and analysis practices are consistent across departments, not rebuilt per team. Our key data is reliable, available and accessible where it is needed. Systems and departments exchange data without manual workarounds. We keep an inventory of current and planned AI use cases. AI governance is anchored in how we deliver projects, not just a policy document. Not yet Partly Yes, clearly EXPLORING Value is still scattered. The exercise gives you one shared picture of where data and AI matter, before any tooling decisions. BUILDING Foundations exist. The gap is turning them into prioritised, department-crossing use cases with one executable roadmap. SCALING You are ready to industrialise. Governance, lifecycle and a scalable data foundation become the differentiators. READINESS CHECK #### How ready is your organisation? Six statements about your business, data and AI practice: a quick first AI readiness assessment. No email required, just an honest signal of where to start. Pragmatic. Strategic insight combined with delivery experience, from vision to action. Cutting-edge. On the frontier of AI, with the discipline to make it work in production. Vendor-neutral. Exploration deliberately decoupled from execution. We build internal capability, not dependency. WHY ARINTI #### From ambition to measurable impact It is important for us to work with an open mind. The Flemish labor market will face new challenges in the next 10 years. We suspected that with all the data we have available, innovative solutions are possible that we hadn't thought of ourselves. Companies and startups that are usually busy with other topics probably have a refreshing perspective. The purpose of the challenge was to encourage them to work out ideas. ## Services ### Engineering We close the gap between ambition and impact by building solutions that are designed to deliver value from the start. That could be a custom AI application, a mesh of AI agents, or an autonomous agent embedded in your workflows. --- ### Data Great AI is only as good as the data behind it. We build the platforms and pipelines that turn scattered, messy data into a structured, reliable foundation, and then put it to work with models that deliver unique insights and AI solutions with measurable outcomes. --- ### Advisory Before building anything, you need to know where AI will actually move the needle. We work with your leadership to cut through the noise, assess what's realistic, and shape a strategy that connects AI to real business outcomes — not just technology ambitions. ## Cases ### [Sinerisk, independent insurance overview app, from idea to launch in two months](https://arinti.ai/cases/sinerisk) Client: Sinerisk · Published 2026-05-15 > Sinerisk is an insurtech app that bundles a household's insurance policies into one plain-language overview. Flagging coverage gaps and overlaps, and helping users prepare a claim file when something goes wrong. Together with founder Mattias Vercauteren, we took it from idea to launch on the App Store and Google Play in just two months. Three core features A personal insurance overview: every policy, every premium, every renewal date, in a single dashboard on the user's phone. An insurance analysis layer that compares contracts side-by-side and surfaces gaps in coverage, overlapping clauses, and differences in limits and deductibles. And a claim preparation flow that helps users chronologically capture facts, photos, and the right policy details so they can hand their insurer a complete file from the first contact. CONTEXT #### An independent layer on top of insurance Most Belgian households hold several insurance contracts spread across different insurers and brokers. Policies arrive by email, sit in broker portals, and pile up in paper folders. Each contract uses its own terminology. Coverage gaps and overlaps between contracts are common but invisible to the policyholder. When something does happen, users lose hours collecting documents and remembering what to send to whom. Sinerisk's founder, [Mattias Vercauteren](https://www.linkedin.com/in/mattiasvercauteren/), wanted to fix this with an independent app: a single place where a user uploads their existing policies and sees, in plain language, what they have, what is covered, what is not, and what to do when something goes wrong. Independence is central to the proposition. [Sinerisk](https://sinerisk.be/) is not a broker, not an agent, and does not sell insurance. The product sits as a neutral layer between the user and their existing insurers. A weekly feedback loop We worked with short iterations of one week. A working app went to the phones of the team at Sinerisk, not a staging URL or slide deck. We used these short iterations to validate functional decisions and to spot product issues early, before they had compounded into rework. The shorter the loop, the cheaper the change. APPROACH #### Scope tightly, iterate with AI in the loop Two months is not a lot of time to take a product from idea to production. The first decision was about scope. We sat down with Sinerisk and built a shared list of every feature and service the product could include. Then we decided together on what was really needed to validate the three core flows with real users. Everything else, broker integrations, premium recommendations, referral programmes was moved to a later release. The second decision was about how to use AI during the build itself. We embedded AI in day-to-day delivery but kept the engineering decisions in the hands of our team. This approach tightened the feedback and development cycle, which is what mattered on a two-month timeline. Flutter for one codebase, two stores The mobile app is built in Flutter. One codebase ships to both the App Store and Google Play, with a single design system and a single test suite. Flutter let the team focus on the user experience instead of on platform parity, and the resulting app performs to native standards on both platforms. Microsoft Azure The backend is .NET running on Azure. Authentication, document storage, policy parsing, and the API layer that the Flutter app talks to all live on managed Azure services. The choice was deliberate: a stack the team knows deeply, with mature security tooling, and the scaling headroom the product needs as it moves from early users to a wider rollout. AI-based policy document parsing The hardest technical problem in the product is reading a user's existing policies. Belgian insurers each format their policies differently: different PDF templates, different terminology, different structure. We built a document parsing pipeline that extracts the structured data Sinerisk needs (policy type, premium, term, limits, deductibles, broker details) from whatever the user uploads: PDFs, screenshots, photos of paper documents. The pipeline was prototyped in week one and iterated weekly against real policies, so the accuracy on launch day reflected real-world variety rather than a clean lab dataset. Security and scalability Insurance data is sensitive personal information. Security was not treated as a follow-up phase. From the first sprint, the backend used managed identities, encrypted storage, scoped access policies, and audit logging. The same applied to scalability. The API and document parsing layer were designed to scale horizontally on Azure from day one, not retrofitted when usage picked up. Metabase reporting From the first release, every flow in the app emits structured events, and every event lands in a Metabase instance that Sinerisk uses daily for monitoring and analytics. The dashboards cover both sides of the product. On the user side: how many policies are being uploaded, which document types parse cleanly and which do not, where users drop off in the claim preparation flow, which insights they read and act on. On the business side: signup conversion, retention by cohort, early-user feedback signals, and the metrics tied to Sinerisk's strategic goals. Sinerisk has end-to-end visibility into product usage and business performance. IMPLEMENTATION #### Flutter and Microsoft Azure The technology choices were made for one reason: ship a quality consumer app, fast, on a stack that can grow with the product without major rewrites later. We had two months to go from a slide deck to a real app in users' hands. Arinti kept the scope honest, what was needed to validate the product, and nothing more. Only two months later we had launched the app on the app stores. RESULTS #### What two months made possible Two months from the first scoping session, Sinerisk was live on both the [App Store](https://apps.apple.com/be/app/sinerisk/id6761456547) and [Google Play](https://play.google.com/store/apps/details?id=be.fairadvantage.sinerisk), with the core features working end to end. For Sinerisk, that timeline is the result: every week not in market is a week not learning and gathering feedback. AI in the delivery loop is what made this timeline possible. AI did not replace engineering decisions; it tightened the cycle between them. --- ### [A hybrid governance model for data and AI at scale](https://arinti.ai/cases/mercedes) Client: Mercedes-Benz Customer Assistance Center (CAC) · Published 2025-06-15 > Mercedes-Benz operates a Customer Assistance Center serving 40+ markets worldwide. We defined a pragmatic data & AI strategy, aligning governance, ownership, and technical architecture for scaling on Databricks and Azure within a SAFe environment. CONTEXT #### A strong foundation, ready to scale The Customer Assistance Center (CAC) of Mercedes-Benz had invested in modern platforms and demonstrated value through AI initiatives like agent assist, multilingual routing, and automated classification. The ambition was to evolve from a reporting-driven operation toward a data- and AI-enabled organisation that could serve as a global aftersales data domain within the Mercedes-Benz Group. To get there, the organisation needed clarity on three fronts: how to structure ownership and governance as the operation scales, how to align the technical architecture with corporate standards, and how to embed data quality and compliance into the way teams work day to day. APPROACH #### Mapping current state, defining the path forward We ran a series of targeted deep-dive sessions with business, IT, operations, and the Intelligence & AI teams. The goal was to map the current state across people, process, and technology — and to identify what needed to change to support the next phase of growth. The engagement covered the full spectrum: data governance, data quality, technical architecture, AI readiness, and compliance (GDPR, EU AI Act). Each domain built on the previous one, with validation workshops to ensure findings reflected operational reality. We delivered a unified strategic alignment canvas for Data & AI, a gap analysis of current versus target state, and a hybrid governance model that combines SAFe delivery practices with Data Mesh principles. The output was not a theoretical framework, but a set of concrete recommendations, quick wins, and a phased implementation roadmap. Strengthen the organisational foundation Define clear ownership structures and governance responsibilities that can scale with the organisation. Ensure that accountability is distributed across domains rather than concentrated in a few roles, and that governance is embedded into how teams plan and deliver. Harmonise the technical architecture Consolidate the data platform into a unified structure with shared standards for ingestion, transformation, lineage, and monitoring. Align the local data catalogue with the corporate metadata platform to ensure end-to-end governance and traceability. Embed governance into the delivery process Make data quality and compliance part of the planning and delivery lifecycle — not something that is bolted on afterwards. Introduce measurement-driven feature approval and automated quality gates within the existing SAFe cadence. RECOMMENDATIONS #### Three priorities for scaling The strategy centred on three priorities that together form the structural foundation for scaling data and AI capabilities across the organisation. Core deliverables We handed over a connected set of artefacts that together give the CAC everything it needs to move from planning to execution. These include a gap analysis across people, process, and technology; an alignment strategy between business and IT with clear ownership structures; a hybrid governance model combining SAFe and Data Mesh principles; a 12–24 month implementation roadmap with quick wins in the first six months; and a compliance readiness assessment for GDPR and the EU AI Act. Together, these give the CAC a clear path to scale its data and AI capabilities, with the structural foundations needed to do so sustainably across 40+ markets. RESULTS #### A strategy ready for implementation The engagement produced a pragmatic strategy and a roadmap ready to put into action. 40+ Markets served by the CAC 12–24 months Implementation roadmap delivered <6 months Time to first quick wins --- ### [Governed Databricks data platform for global operations](https://arinti.ai/cases/volvo-logistics) Client: Volvo Logistics · Published 2025-04-15 > Volvo Logistics distributes over 700,000 spare parts worldwide. We built a governed Databricks data platform that unified operational data from six global sites, delivering up to 40% efficiency gains, 99% reduction in pipeline latency, and self-service analytics for business users. The structural gap The visible problem was fragmented reporting. The underlying issue: Volvo Logistics had no single, governed data layer that could serve all domains with consistent definitions and controlled access. Years of mergers and acquisitions had left systems that had never been designed to talk to each other. Inventory planning occurred warehouse by warehouse, with limited visibility into what resources could be shared. What was needed was a platform that could ingest from every source system, standardise definitions across domains, enforce governance at every layer, and still move fast enough for operational decision-making. We scoped and built that platform together with the SML team. CHALLENGE #### Aftermarket logistics at global scale The right part needs to reach the right workshop, fast. Volvo Logistics' SML division is responsible for getting spare parts from suppliers to dealers and workshops worldwide. It is a high-volume, time-sensitive operation where planning, warehousing, and transport decisions all depend on accurate, current data. Before this project, that data was spread across dozens of source systems. Mainframes held operational records. SAP held transactional data. Middleware tools like IBM CDC and IBM MQ captured changes from source systems. Azure Event Hub carried live signals from equipment and processes. Each domain — Warehousing, Transport, Planning — maintained its own extracts and its own version of the numbers. Scheduled pipelines ran alongside live data feeds, uncoordinated and often duplicating effort. Governance had been applied inconsistently. When business users needed a new report, the request went into an IT queue and could take weeks to surface. Some requests did not surface at all. Bronze: raw data, preserved as-is Raw data from every source system lands in the Bronze layer unmodified: mainframe dumps, SAP extracts, change feeds from middleware, live streams from Event Hub. Nothing is filtered or altered. This creates an auditable record of everything entering the platform, regardless of format or origin. Silver: standardised and enriched A single ingestion framework handles the transformation work across all source systems. It is driven by configuration rather than custom code: adapting when data structures change, enforcing data types and validation rules, and tracking how records evolve over time. Mapping tables standardise definitions so that a shipment means the same thing whether it originates in Warehousing or Transport. Gold: business-ready data products Curated datasets serve Power BI dashboards, operational applications, and Databricks' AI/BI tools. Each dataset is organised per domain and governed through Unity Catalog, so Warehousing sees its data products and Transport sees its own. Access is controlled down to the level of individual dataset structures. PLATFORM #### SMLCloud: the Databricks data platform SMLCloud is a large-scale Azure Databricks platform that unifies all of Volvo Logistics' operational, warehouse, and supply chain data into governed, analytics-ready data products. The architecture follows a medallion pattern — a layered approach where data gets progressively cleaner as it moves from raw ingestion to business-ready outputs. GOVERNANCE #### Governance built into every pipeline Unity Catalog manages every dataset, every report view, and every access permission on the platform. We migrated more than 30 domain schemas from the previous catalogue system (Hive Metastore), standardised access controls, and ensured that all automated processes run under controlled, auditable identities rather than personal accounts. Access is role-based: a transport analyst sees transport data, not warehouse financials. Data lineage is tracked end to end — any number can be traced back to its source. Governance here is not a policy document. It is built into the platform's daily operation: every pipeline, every deployment, every access request runs through it. Streaming pipelines We introduced a real-time track alongside the scheduled pipelines. Machines send status updates via MQTT through Azure Event Grid. Each device authenticates with its own certificate, so the data stream is both secure and traceable. Spark Structured Streaming transforms the incoming signals on arrival. A lightweight database holds the latest reading per metric, and an automated function pushes it to Salesforce for rental staff worldwide. Field engineers and rental managers can now check live machine status from their phones. Lakeflow migration The platform is migrating its scheduled pipelines to Lakeflow Declarative Pipelines, a newer Databricks framework that removes the need for intermediate storage between processing steps — resulting in up to 40% efficiency gains on routine data tasks and a 99% reduction in pipeline latency. REAL-TIME #### From batch-only to near-real-time The original data flows ran on a schedule: nightly jobs that processed the previous day's records. For a logistics operation where decisions depend on what is happening now, that one-day lag was a constraint the business worked around rather than accepted. With just a simple click, we can transition our metadata scheduling from triggered to continuous and back again, without the need to dive into complex code. Spark Declarative Pipelines can handle schema changes almost automatically. It's a nice feature. SELF-SERVICE #### Data without an IT ticket We built Databricks' AI/BI Genie Spaces as the interaction layer for business users. It lets people ask questions about their data in plain language and get answers drawn from the curated Gold datasets. Power BI dashboards serve structured reporting needs. Genie Spaces serves the ad hoc ones: the questions that come up in a meeting and used to require a data analyst to answer. For a growing share of the organisation's data needs, the queue between question and answer has been cut from weeks to seconds. DELIVERY #### Delivery at scale across six sites With hundreds of automated data jobs across dozens of domain packages, the platform needed rigorous deployment discipline. We industrialised delivery using Databricks Asset Bundles and Azure DevOps CI/CD. Every change follows a controlled path: development, quality assurance, pre-production, production. Only the packages that actually changed get deployed. Automated tests verify logic before anything touches a live environment. Code-format checks enforce consistency across the engineering team. The platform runs across six global sites, processes data from over 200 source tables, and supports the daily operations of a logistics network that serves Volvo's aftermarket worldwide. It has been running since May 2022 and continues to expand. Today, we get an integrated view of where spare parts are, the value of spare parts across warehouses and the potential costs involved in shipping parts from warehouse to warehouse or dealers. 40% Efficiency gains on routine data tasks 99% Pipeline latency reduction 700K+ Spare parts distributed worldwide 6 Global sites on one platform 30+ Domain schemas governed 200+ Source tables processed RESULTS #### What this made possible Volvo Logistics started this project with a data environment that collected data but could not connect it. Four years in, the platform has evolved from batch-only processing to a hybrid architecture combining batch and real-time pipelines on one governed Databricks platform. The next steps are already in scope: predictive maintenance based on sensor patterns, data-as-a-service offerings where machine insights become a product, and continued migration to Lakeflow Declarative Pipelines for fully real-time ingestion. --- ### [Forecasting sales across 14 product categories](https://arinti.ai/cases/unilever) Client: Unilever · Published 2024-06-01 > Sales forecasting for Unilever Belgium, predicting daily and monthly sales volumes across 14 product categories using category-specific machine learning models, with Power BI dashboards for the commercial team. CHALLENGE #### Forecasting across 14 product categories Unilever Belgium wanted to forecast sales at product level across 14 categories, with two distinct time horizons: daily predictions for the next two weeks (for operational planning) and monthly predictions for the next year (for strategic planning). The existing process lacked the granularity and automation the commercial team needed. Gaps and inconsistencies The dataset contained numerous missing values, skewed distributions, and limited feature information. Some products lacked sufficient historical data entirely. We filtered those out, applied interpolation and mean-value filling for numeric gaps, and introduced an "unknown" category for missing categorical features. Feature engineering We created lag features from previous sales periods to capture trends, and identified seasonal patterns in the data — for example, certain product categories show consistent spikes during summer months. For monthly forecasting, the model looks back two years to pick up seasonal signals from previous cycles. DATA #### Working with imperfect data The available sales data presented several challenges that had to be solved before any model could be trained. One model per category Consumer behaviour varies significantly across product categories, so we built a separate model for each of the 14 categories rather than forcing a single model to generalise across all of them. This gave each model the room to learn category-specific patterns. Automated retraining We built automated scripts that detect new categories, train models, and generate forecasts — so the system scales without manual intervention when the product portfolio changes. APPROACH #### Category-specific models We framed the problem as a regression task and tested several model architectures: LSTM neural networks, tree-based models, and gradient boosting. RESULTS #### What the models delivered The best-performing models achieved an RMSE of 0.6 and an r² of 0.70, meaning the models explained roughly 70% of the variance in actual sales, which is strong for product-level retail sales, where consumer noise typically caps explained variance well below the headline numbers seen in cleaner forecasting domains. Results were delivered through Power BI dashboards designed for the commercial team, not data scientists. --- ### [AI scheduling tool wins back 17 production days per year](https://arinti.ai/cases/griffith-foods) Client: Griffith Foods · Published 2024-01-15 > AI scheduling tool wins back 17 production days per year: an algorithm that optimises production line cleaning sequences at Griffith Foods' Herentals facility, reducing wet cleanings by 5%, saving 1,000m³ of water annually, with payback in 10 months and global rollout to all 21 sites. The manual planning burden The production planners knew that some product sequences could safely run with a dry clean in between. Two yellow batches with the same allergen profile don't need a full wet wash. But calculating which sequences were safe, factoring in colour, allergens, texture, odour, and ingredient composition, took hours of manual work. And the planning changed every day. PROBLEM #### 165 lost production days, hiding in plain sight Griffith Foods is an American developer and producer of customised food ingredients, operating 21 production sites worldwide. At its Herentals facility in Belgium, rising customer demand was outpacing available production capacity. The answer sat in the cleaning schedule. Between every production batch, lines underwent a full wet cleaning — water, detergent, 30 to 90 minutes of downtime. Across all lines at the Herentals site, that added up to roughly 165 lost production days per year. Arinti did not just dump a lot of technology into our organisation. They listened to our demands and developed a tool that works for us. Our planners find the tool easy to use. First attempt: recipe attributes Our first approach used a model built on recipe-level attributes: allergens, odour, taste, and colour classifications per product. Several of these parameters resisted consistent quantification. Odour and taste, in particular, could not be encoded cleanly enough for an algorithm to act on. The pivot to ingredient data We switched to the ingredient lists stored in Griffith Foods' ERP system. Ingredients are factual, granular, and already maintained as structured data. Two products sharing the same base ingredients are likely compatible. The ERP data gave the algorithm a foundation it could work with reliably. APPROACH #### Why the sequencing problem was harder than it looked Griffith Foods produces hundreds of distinct recipes, each differing in colour, odour, texture, and allergen content. The rules governing when a dry clean suffices and when a wet clean is mandatory depend on the specific pair of products running back-to-back. A tool the planners actually use The algorithm runs behind a purpose-built interface designed for the production planners at Herentals. Where the previous process took hours of manual calculation each day, the tool generates an optimised schedule in minutes. Planners can also simulate what happens when an urgent order needs to be inserted mid-schedule. A companion app gives floor operators visibility into which batches run on which machines. BUILD #### Building the scheduling algorithm The algorithm works by matching products based on shared characteristics derived from their ingredient lists. Products that overlap in allergen profile, colour, and composition can follow each other with a dry clean. Products that diverge on any critical parameter require a wet clean. From these pairwise compatibility scores, the algorithm builds sequences: chains of products that can run consecutively with only dry cleans between them. The sequences are then stitched together with wet cleans at the boundaries. From one factory to twenty-one The results at Herentals triggered a decision to roll the tool out globally. Two additional European factories came next, with the explicit goal of using their operational data to refine the algorithm further. The architecture supports this: the algorithm's logic is configuration-driven. Ultimately, all 21 Griffith Foods production sites worldwide are scheduled for deployment. RESULTS #### What changed on the factory floor The Herentals site saw results soon after deployment. The production team reduced wet cleanings by 5%, with a target of 10% or more. That translated into 17 production days won back per year at a single site. Water consumption fell by roughly 1,000 cubic metres per year, with a corresponding drop in detergent use. The project reached payback in 10 months. For the planners, a daily planning task that had consumed hours became a few minutes of review and adjustment. --- ### [GenAI-powered virtual assistant for citizen services](https://arinti.ai/cases/stad-kortrijk) Client: Stad Kortrijk · Published 2023-10-01 > AI-powered virtual assistant for citizen services - the first GenAI-powered chatbot on a Flemish municipal website, combining conversational AI with semantic search to answer citizen questions in plain Dutch, directly from the city's own content. What the city actually needed What the city needed was a virtual assistant that could understand natural-language questions in Dutch, find relevant answers from its own content, and guide citizens to the right outcome without requiring them to navigate the site at all. CONTEXT #### A city with digital ambition and a practical problem Kortrijk had invested in its digital infrastructure. A well-maintained Drupal website, citizen portals, links to Flemish government services. But the fundamental citizen experience had not changed: residents still searched, scrolled, and more often than not, picked up the phone. The city's contact centre handled a high volume of questions that already had answers somewhere on the website. Opening hours, permit procedures, waste collection schedules, event information. All published. All hard to find if you did not already know where to look. CMS lock-in and the limits of keyword search The content lived in Drupal. Any solution needed to plug into the city's existing CMS without requiring a parallel content management workflow. And the assistant had to go beyond keyword matching. Citizens do not ask questions the way a search engine expects them to. 'When is city hall open?' and 'Can I go to the municipality tomorrow?' are the same question, but a traditional search treats them as entirely different queries. The trust bar for government content Then there was the constraint that matters most in government: trust. Answers had to be accurate, traceable, and grounded in the city's own published content. Hallucinated responses were not an option. We scoped the project together with Kortrijk's team around that principle from day one. CHALLENGE #### What made this harder than it sounds Two constraints shaped the approach from the start. Intent over keywords When someone asks a question, the assistant does not match words. It interprets meaning, retrieves the best-matching content from indexed sources, and generates a concise answer with a direct link to the source page. Every response is grounded in published content. Built into Drupal, not bolted on The assistant was built as a custom Drupal module, integrated directly into the city's existing website. It appears as a chat interface and as a semantic search layer on the search page, replacing the standard Drupal search with something that actually understands what people are asking. One codebase, one set of content, no parallel systems to maintain. APPROACH #### Conversational AI meets semantic search We built a virtual assistant that combines conversational AI with hybrid search, merging keyword-based retrieval with semantic search technology. The system indexes the city's web content and documents, then uses Azure AI Services to interpret the intent behind a citizen's question and surface the most relevant answer. Deeplink integrations and personalisation Tourism content from a separate domain was integrated into the main search. Deeplink integrations with eGovFlow now route citizens straight into municipal e-services: appointment booking, permit applications, complaint forms. From page retrieval to generated answers Early versions used pure semantic matching to retrieve relevant pages. Later iterations introduced OpenAI-powered response generation: the model reads the retrieved content and composes a direct answer. We regularly benchmark newer model versions against accuracy and cost, upgrading when the numbers justify it. Analytics dashboard A PowerBI dashboard gives the city's team visibility into what citizens are actually asking, where the assistant performs well, and where content gaps remain. SOLUTION #### Expanding across the city's digital services The core system does two things. It handles citizen questions through a conversational chat interface, and it replaces the default Drupal site search with a semantic search layer that returns AI-generated answers alongside traditional page results. Both draw from the same indexed content. Both stay current without manual intervention. RESULTS #### What Kortrijk can do now Kortrijk was the first city in Flanders to deploy a GenAI-powered virtual assistant on its municipal website. VRT NWS covered the launch. Citizens ask questions in plain Dutch and get direct, sourced answers. Around the clock. The assistant handles queries about waste collection, permits, opening hours, event information, and dozens of other topics that previously required a phone call or a visit to the contact centre. Because everything runs through Drupal, the editorial team maintains one set of content. When a page is updated, the assistant's answers update with it. No drift between what the website and what the bot says. --- ### [Automating parcel label extraction](https://arinti.ai/cases/latvian-post) Client: Latvian Post · Published 2022-12-15 > Automating parcel label extraction at Latvian Post, using Azure Form Recognizer and computer vision to liberate 2,000+ hours of manual work per month by automatically reading and completing missing parcel data. CONTEXT #### A queue that never emptied Latvian Post is the national postal operator of Latvia, handling parcel volumes that scale with e-commerce demand across the Baltic region. The organisation had already shown appetite for innovation, and was among the first European postal services to trial delivery robots. Its internal data processing, however, had not kept pace. Parcels with incomplete sender or recipient data were routed to a manual review queue. There, a dedicated team of 20 full-time employees inspected each package, identified missing fields, and entered the data by hand before the parcel could continue through the system. At scale, this queue never emptied. The engagement came through Future Hub, a Scandinavian accelerator running its second wave of the Open Innovation Sustainability programme. The brief was to find a technical solution to a sustainability challenge that the host organisations faced in daily operations. Latvian Post's manual data entry problem was exactly the kind of high-volume, repeatable burden that automation could address. To carefully extract missing data from the parcels we process, we currently employ a team of 20 full-time employees, who definitely have better things to do. CHALLENGE #### Precise, at production speed The goal was precise. Use image processing on the parcels themselves to extract missing data automatically, validate what was already in the system, and route each package into a managed review cycle so the algorithm either got it right or learned from its mistakes. Automatically, at production speed, inside the existing workflow. Speed mattered as much as accuracy. A proof-of-concept that worked cleanly in a lab but stalled under real volume would be of no value. The solution had to handle the natural variation in label formats, fonts, and print quality that characterises real-world parcel data, and hold its performance at scale. The whole programme, from first handshake to demo day, lasted ten weeks. IMPLEMENTATION #### Azure Form Recognizer, trained on real data As a Microsoft Partner, we know what Azure can do — in this case, Azure Form Recognizer. It's a building block that already extracts millions of data points from images every year, which made it the right starting point for the problem at hand. We got to retraining the model straight away, and built the API alongside it. Sending images in and getting clean, labelled data back was as much of a challenge as the data extraction itself. This was the kind of work our data scientists and data engineers do best, and we had a testable solution within weeks. That gave Latvian Post room to send us more image data, and gave us the time to push the recognition speed down to 10 seconds on average. Almost eight times faster than the same task done manually. We only hit that speed by running on Azure Cloud, where Microsoft's GDPR commitments gave us the compliance footing we needed. > 85% Extraction success rate — data extracted, validated, and entered without human involvement 2,000 hrs Freed per month at production volume 8× Faster than manual processing — average extraction time of 10 seconds per parcel #### Results at demo day RESULTS #### Time freed, data quality improved With more than 85% of parcels processed automatically, the team previously assigned to manual data entry could be redirected to higher-value work. The roadmap at close of the engagement included deeper integration into Latvian Post's approval workflows, a threshold-based automatic release for high-confidence extractions, migration to an on-premise deployment, and a planned handover to the internal IT team to own and extend the system. --- ### [Smooved - Recommending brokers to house owners](https://arinti.ai/cases/smooved-broker-recommendation) Client: Smooved · Published 2022-07-22 > Belgium has over 5,000 active brokers and no reliable way for property owners to compare them. We built the recommendation engine behind HouseMatch.be, analysing broker data daily using an Azure data warehouse, BERT sentiment analysis, and a custom explainable matching model. Why explainability was a requirement Smooved needed a recommendation engine that could suggest the ideal broker based on real performance data, not just proximity. The system had to be explainable to both end-users and Smooved's sales team, who needed transparent rankings to keep brokers on board. CONTEXT #### The problem with finding a broker Three quarters of people who sell a home describe the process as stressful. With thousands of reviews scattered across Google, Facebook, and other sources, no reliable tool existed to cut through that volume and surface the right match for a specific property in a specific area. Smooved built HouseMatch to solve that problem. What they needed was a recommendation engine grounded in real, up-to-date data, with decisions that could be clearly explained to the brokers it ranked. Without explainability, the system couldn't be sold. Without the data, it couldn't be trusted. What the first attempt taught us We started where most data projects start: collecting data. The Belgian broker market was mapped, review sources were ingested, and several machine learning models were trained against the resulting dataset. They fell apart under real conditions: the outputs weren't explainable. When Smooved's team asked why a broker ranked where they did, the models had no useful answer. The consequence was predictable: brokers didn't trust the system, and getting them on board became very difficult. Work stopped. What had seemed like productive momentum was pointed in the wrong direction. So we sat down again, this time with a genuinely open brief, and rebuilt the problem statement from scratch. Defining what the system actually had to do The second approach was built on a firmer premise. Every broker score had to trace back to specific, interpretable factors: the kind that could survive a direct sales conversation. From that requirement, we worked backwards with Smooved: what data was genuinely necessary, what introduced noise, what the model had to explain and to whom. Each sprint involved workshops where both teams stress-tested outputs together, working through edge cases until the recommendations held up under scrutiny. Where the first attempt had moved fast and assumed alignment, the second confirmed it at each step. APPROACH #### A credible concept with a hard execution problem Smooved had a clear product ambition: HouseMatch would let property sellers scan their area and receive ranked broker recommendations based on objective, verifiable data. The concept was credible. The execution required building something that didn't exist: a structured, continuously updated view of the Belgian broker market, drawn from fragmented and inconsistent sources. The recommendations also had to be defensible. A broker who saw a low ranking would ask why. Smooved's sales team needed to be able to answer that question directly, without retreating to vague model outputs. Explainability wasn't a nice-to-have. It was a commercial requirement. The data infrastructure A four-layer Azure data warehouse forms the foundation. Raw data from Smooved's internal APIs sits alongside broker profiles and reviews ingested from Google and Facebook Places through external API integrations. Each layer refines and enriches the data below it. By the gold layer, every broker has a coherent, continuously updated profile drawn from multiple independent sources. At scale, that means over 5,000 brokers analysed every day and more than 120,000 reviews processed continuously, with no manual intervention. Sentiment analysis at scale Sentiment analysis across broker reviews runs through a Zero-Shot Learning BERT model. Zero-shot was the right choice for this problem: it required no labelled training data specific to the Belgian real estate domain, which removed months of preparation time and avoided the bottleneck of building manual sample sets. The model reaches 95% accuracy on sentiment classification. The output feeds directly into broker profiles: services where a broker consistently performs well, and areas where they fall short, identified automatically and updated continuously. The recommendation model The recommendation itself is produced by a custom mathematical model, designed from the start to be interpretable. Each output is traceable. When a broker ranks highly or poorly, the factors driving that position are visible to Smooved's team and, where relevant, to the broker directly. The model is exposed through a versioned API endpoint, which allows us to run concurrent versions in parallel, compare performance, and promote changes to production in a controlled way. The production version went live on schedule, ready for HouseMatch's launch. IMPLEMENTATION #### What was built 5,000+ Brokers profiled daily 120,000+ Reviews ingested continuously ~25% Of the Belgian broker market reached on launch day, generating 100+ new partnership requests #### Impact at launch RESULTS #### What happened at launch HouseMatch launched. On day one, the platform generated visits to nearly a quarter of all active Belgian brokers. More than 100 additional brokers reached out to Smooved to request a partnership. The ongoing pipeline processes 5,000+ brokers and 120,000+ reviews daily to keep rankings current. With HouseMatch, we want to help homeowners to navigate their real estate transaction to the happiest possible outcome. At the same time, we want to offer real estate brokers more visibility with sellers and landlords. --- ### [From SharePoint archive to knowledge tool](https://arinti.ai/cases/fit-sharepoint-knowledge-tool) Client: Flanders Investment & Trade · Published 2022-01-15 > AI-powered enterprise search for Flanders Investment & Trade, turning 15,000 documents into a searchable knowledge base with automatic metadata extraction, image analysis, and semantic search on Azure AI Search, supporting 10,000 client enquiries per year. CONTEXT #### An agency built on knowledge Flanders Investment & Trade is the Flemish government agency responsible for supporting international business. From its headquarters in Brussels and 70 offices worldwide, FIT advises companies on export opportunities, conducts market research across sectors, and tracks global economic trends. The agency handles 10,000 client enquiries a year. Every one of those enquiries depends on experts who can access the right knowledge quickly. That knowledge is extensive. FIT's analysts produce detailed reports on sectors from food and agri to quantum computing. Over time, this built into a SharePoint archive of 15,000 documents — with no sign of slowing down. Manual categorisation overhead Every document entering the archive had to be tagged and filed by hand. 'Structuring, categorising and archiving data is a very labour-intensive affair,' said Yves Ruland, Data Project Manager at FIT. The time spent maintaining the structure was time not spent on analysis. Knowledge silos across the organisation FIT's knowledge was distributed across departments, geographies, and years. There was no reliable way to surface connections between pieces of information, or to find expertise buried in a document from a different team and a different year. CHALLENGE #### The limits of folders and keywords FIT's existing setup relied on folder structures and keyword matching. Both required the user to already know roughly where to look and what to call it. For a new analyst, or an expert working outside their usual domain, that was a genuine barrier. Arinti stood out thanks to their pragmatic approach. Beyond text: image and entity recognition The extraction goes beyond text. The system identifies locations and company names mentioned in documents, and analyses image content to generate additional metadata. A photo taken at a trade mission — with no caption — can still become a searchable, tagged asset in the archive. Hybrid search: keywords meets semantics Search works as a hybrid: keyword matching combined with semantic understanding. A query doesn't need to use the exact terms in a document to surface it. The system reads intent, then filters results using the auto-generated tag layer. Metadata-driven, not folder-driven The classic folder tree gave way to a metadata-driven approach. Documents no longer need to live in the right folder. They need good tags, and the system handles the tagging. 'Everything is so fleeting now; if you structure things too much, the future catches up with you,' said Ruland. IMPLEMENTATION #### AI-powered enterprise search on Azure Arinti built FIT Enterprise Search on Microsoft Azure AI Search, integrated directly with FIT's existing SharePoint environment. The core of the system is automatic metadata extraction: when a document enters the archive, the AI analyses its content and generates tags without any manual input. RESULTS #### From search engine to knowledge tool Bulckaert calls FIT Enterprise Search a knowledge tool, not just a faster search engine. Experts preparing for a client meeting can pull relevant market data in seconds. 'FIT Search acts as a virtual, extra brain,' said Bulckaert. The graphic search results — which surface connections between related documents — also enabled discovery. Experts found relevant material from different teams or previous years that they hadn't known to look for. Automatic tagging reduced the manual overhead of categorisation. New documents could be added faster, with less friction. 'The improved search function means we can now add to our research faster and more intelligently,' said Ruland. With Arinti we took a close look at our knowledge and business databases.The company developed an intelligent search robot that automatically uncovers patterns, connections and derives information within the many thousands of FIT documents and data related to export and foreign investments. In this way, our employees can make knowledge and expertise from within and outside FIT available to the business world even faster. WHAT'S NEXT #### Building on the foundation At close of the engagement, FIT's roadmap included personalised marketing and lead generation as the next AI focus areas. The search tool had demonstrated that machine learning was a practical option inside FIT's infrastructure — not just a concept. --- ### [DAIQUIRI - AI for live sports reporting](https://arinti.ai/cases/daiquiri) Client: imec.icon Research Consortium · Published 2021-09-30 > DAIQUIRI (Data & AI for Quantified Reporting In sport) was an imec.icon research project that developed AI algorithms and a sensor data platform for enriching live sports broadcasts with real-time athlete data, demonstrated in professional cyclocross and elite hockey. CONTEXT #### Sensor data that never reached the audience During professional sports events, athletes generate large volumes of data through wearables and sensors — heart rate, power output, speed, positioning. This data was used primarily for coaching and performance analysis. It rarely reached the broadcast or the audience. The gap was not in the data itself but in the translation: turning raw sensor streams into narrative elements that commentators, content editors, and viewers could use in real time. No platform existed to bridge that gap at production speed. PROJECT #### An imec.icon research consortium DAIQUIRI was an imec.icon research project funded by imec and Agentschap Innoveren & Ondernemen, running from October 2019 to September 2021. The consortium brought together experts in sports event capturing, sensor data platforms, editorial tools, and interactive user experiences. The project had four defined outcomes. First, optimising sensor connection and data quality to reduce data stream sizes by 30%. Second, enabling insight-driven data capturing with real-time dynamic visualisation. Third, developing AI algorithms for generating different types of story snippets from sensor data. Fourth, unlocking storytelling techniques that complement traditional reporting. APPROACH #### From sensor stream to story snippet The core technical challenge was building a scalable data workflow that could ingest IoT data from athletes and equipment, process it through AI enrichment, and output structured narrative elements — all within the latency constraints of live broadcasting. The AI layer handled four challenges. Data overload meant filtering signal from noise in high-frequency sensor streams. Sensor-video matching synchronised data points with broadcast footage. Dynamic captioning generated contextual text overlays. Multi-modal story generation combined data, video, and text into coherent segments for different output formats. DEMONSTRATION #### Cyclocross and elite hockey The platform was demonstrated in two live sports scenarios: professional cyclocross and elite hockey. Sensor data and insights gathered by the platform were used to create templates for real-time visualisation, fed into a content creation dashboard. Media professionals used the dashboard to dynamically add data-driven insights to their coverage. --- ### [YANA - Chatbot for Youth Safety in Mechelen](https://arinti.ai/cases/yana) Client: Stad Mechelen · Published 2021-07-01 > Chatbot YANA ("You Are Never Alone") guides youngsters safely through Mechelen's nightlife: a Facebook Messenger bot that tracks your journey, confirms safe arrival, and alerts your contacts when something seems off. How it works You provide your departure location, destination, and means of transport. YANA calculates when you should arrive. Once there, it asks you to confirm — and sends a message to your contact person. If there's no response within the timeframe, your contact is automatically called via Proximus so they can reach out to you. CONTEXT #### Chatbot YANA guides youngsters safely through Mechelen YANA — You Are Never Alone — is a chatbot on Facebook Messenger that acts as a digital nightlife assistant. When going out, youngsters chat with YANA like they would with a friend. It's super user-friendly and useful for when you feel unsafe, especially late at night. Designed for the target audience The chatbot is tailored to youngsters with short answer options and quick replies. Testers confirmed it feels natural and accessible. Testers like Lisa pointed out that parents always ask you to let them know when you've arrived safely, and you often forget. YANA takes care of that automatically. Reporting to the city YANA also allows young people to report issues — violence, damage to public property, or feeling unsafe in a specific area. Reports go directly to the relevant city services, and in case of aggressive behaviour or violence, the police receive an immediate notification. SENSE OF SECURITY #### An extra layer of security for young people Going home alone at night shouldn't feel like a risk. YANA provides a low-threshold digital safety net, no app download needed, just a Facebook Messenger conversation. There are already cities and municipalities with a chatbot, but nowhere is there such a focus on the feeling of safety of young people. By adding a comprehensive reporting function, we can also encourage young people to report more matters. Privacy by design Young people are a vulnerable group, so extra care was taken. The chatbot runs on Microsoft technology for robust security. Users decide how much information to share — when reporting to the city, providing your name or contact info is entirely optional, and only specific city services can access report data. Proximus integration For the SMS and call functionality, we partnered with Proximus to enable automated notifications and phone calls to emergency contacts. UNDER THE HOOD #### Built with Microsoft technology and privacy-first YANA was built using the Microsoft Bot Framework and integrates with Proximus for SMS and call functionalities. Privacy was a top priority throughout development. RESULTS #### Slim in de Stad prize YANA was realised thanks to winning the 'Slim in de Stad' (Smart in the City) prize. The chatbot was publicly launched in July 2021 after testing by ten young users. It's accessible via the Facebook page 'Yana van 2800leeft'. --- ### [TAFUTA - A cognitive search platform](https://arinti.ai/cases/tafuta-cognitive-search-platform) Client: Multiple (VITO, FIT, VLAIO, VIB) · Published 2020-11-01 > Tafuta is our AI-powered Cognitive Search solution. Extracting, structuring, and surfacing knowledge from any document type across your organization using NLP, computer vision, and custom entity detection. Research paper discovery VITO uses Tafuta to offer technological researchers a tool to more accurately retrieve relevant research papers and documents. The data source spans PDFs, Word documents, PowerPoints, and Excel files from on-premise storage, plus external sources like Google Scholar. A custom entity detection skill enables searches based on economical, chemical, and process parameters. Rich result presentation Search results appear in a table showing the most relevant documents first, plus a graph representation showing how documents are connected and related to specific entities (people, organisations) or topics. The document database is automatically populated with new content based on research trends. Custom per deployment Each deployment includes custom models and classifiers tuned to the client's specific business goals and industry requirements. CONTEXT #### Make internal content discoverable with Tafuta Organisations like Flanders Investment & Trade, VITO, VLAIO, and VIB have deployed Tafuta to make their internal content discoverable. Tafuta, Swahili for 'search', is a cognitive search platform that combines full-text search with AI-driven enrichment to turn raw documents into searchable, structured information. It won the Smart Investigation challenge in the 2020 Smart Policing Hackathon. How it works The solution uses NLP and AI services across vision, language, and speech, including OCR, translation, key phrase extraction, and entity detection. Together, these transform raw, unstructured information into searchable content. It handles Microsoft Word, PowerPoint, Excel, PDF, PNG, RTF, JSON, HTML, and XML formats. Indexing and querying The two primary workloads are indexing and querying. Indexing brings text into the solution and makes it searchable by processing inbound text into tokens stored in inverted indexes. Once populated, you can send query requests with relevance tuning, autocomplete, synonym matching, fuzzy matching, pattern matching, filtering, and sorting. HOW IT WORKS #### What's a cognitive search solution? Tafuta is built on the cognitive search pattern: a knowledge retrieval service with built-in AI capabilities. It provides a full-text search engine, persistent storage of search indexes, and integrated AI used during indexing to extract more text and structure from documents. Natural language processing skills Entity recognition, language detection, key phrase extraction, text manipulation, sentiment detection, and PII detection. With these skills, unstructured text is mapped as searchable and filterable fields in the index. Image processing skills Optical Character Recognition (OCR), facial detection, image interpretation, image recognition for famous people and landmarks, and attribute detection like image orientation. These skills create text representations of image content, making visual information searchable. Practical scenarios Scanned documents (JPEG) made full-text searchable via OCR. PDFs with combined image and text where NLP processing produces better results than standard indexing. Multi-lingual content with automatic language detection and translation. Multimedia analysis on audio, video, and images. AI ENRICHMENT #### AI-enriched search engine AI extracts text from images, blobs, and unstructured data sources, making content more searchable. Enrichment and extraction are implemented using cognitive skills attached to the indexer-driven pipeline — both built-in and custom skills we create for your domain. Unstructured content at scale When your raw content is largely undifferentiated text, images, or application files, the search solution identifies and extracts text during indexing, creates structure, and generates new information such as translated text or detected entities. Custom text analytics When your content needs linguistic or custom text analysis, analysers can be configured for specialised processing: filtering out diacritics, recognising patterns in strings, or building domain-specific entity classifiers for finance, science, or medicine. WHEN TO USE IT #### When should you use a cognitive search solution? A cognitive search solution fits organisations that need an in-company search experience similar to commercial web search engines, or need to consolidate heterogeneous content types into a private, user-defined search index. 10+ Document formats supported including Word, PDF, PowerPoint, Excel, PNG, and more NLP + CV Combined natural language and computer vision AI skills for deep content extraction Custom Domain-specific entity detection and document classifiers tailored to your industry #### Platform capabilities --- ### [YouFlanders - Safe tourism during COVID-19](https://arinti.ai/cases/youflanders) Client: Toerisme Vlaanderen · Published 2020-06-08 > Mobile app for Tourism Flanders that provided real-time crowd levels, local COVID-19 measures, and tourism recommendations, built in under three months with Cronos Public Services, Monkeyshot, and Studio Hyperdrive, launched on 8 June 2020. CONTEXT #### A sector looking for answers In March 2020, Belgium's tourism sector came to a full stop. Tourism Flanders surveyed hundreds of operators and stakeholders to understand what was needed for a safe restart. Over 300 ideas and suggestions came back, and the common thread was clear: operators and visitors alike needed real-time information about crowd levels, local measures, and what was permitted where. Tourism Flanders turned this into a concrete brief: a mobile app that connects, informs, and inspires — giving both visitors and operators access to the same up-to-date information. We had no time to lose, but everything had to be done right from the start. Cronos Public Services provided the solution, and we could quickly partner with their competence centres for design, development, and data architecture. What the app did The app showed real-time crowd levels at tourist attractions, local COVID-19 measures per region, and recommendations for nearby destinations. Users could actively report how busy a location was, helping others decide whether a visit was worthwhile. Operators could indicate which measures applied at their venue. Accessibility and privacy The app was built for all users, including people with disabilities. Privacy was a core design constraint — user data was only used for its stated purpose. IMPLEMENTATION #### Three months from concept to launch Tourism Flanders partnered with Cronos Public Services, bringing together Monkeyshot (design), Studio Hyperdrive (development), and Arinti (data analysis and data architecture). The deadline was fixed: the app had to go live on 8 June 2020. SCALING FURTHER #### What came next After the initial launch, the team continued developing the platform. Automated heatmaps were added, showing crowd density based on mobile tower data. The content scope expanded beyond tourist attractions to include natural areas. The data collected through the app also created a foundation for future applications — aggregated crowd and mobility data that Tourism Flanders could use for broader planning. --- ### [COVID-19 triage at pandemic speed](https://arinti.ai/cases/covid-healthbot) Client: Belgian Hospitals (ASZ, AZ Delta, AZ Vesalius, ZOL, AZ Rivierenland, ZH Waregem, SLBO) · Published 2020-04-06 > In March 2020, Microsoft selected Arinti as the sole Belgian implementation partner for COVID-19 Healthbot deployments. Within 12 days, we deployed symptom-checking chatbots at 6 Belgian hospitals, and followed up with a patient triage bot integrated with electronic patient dossiers. CONTEXT #### Hospitals needed triage, fast In early 2020, Belgian hospitals were fielding a surge of questions from concerned citizens about possible COVID-19 symptoms. They needed a way to triage remotely, helping people assess their own risk and receive appropriate advice before arriving at the hospital. At the same time, hospitals had no structured data on how many people were experiencing symptoms, where they were located, or how many might need admission. Microsoft had created a set of Healthbot templates for exactly this scenario: configurable chatbots that could be adapted to each hospital's needs and deployed within days. They needed a Belgian implementation partner with hands-on experience in the Bot Framework. APPROACH #### Selected as Microsoft's Belgian implementation partner Arinti was selected as the sole Belgian partner for the Healthbot rollout, based on our track record with the Microsoft Bot Framework and our seat on the Bot Framework Partner Advisory Board. The Healthbot templates provided the foundation. We customised each deployment to the hospital's specific requirements, integrated QnA Maker for extending the chatbot's response scenarios, and built the front-end UX for embedding on hospital websites. The interface was stripped to essentials — button-driven input, drop-down menus for detail, designed for any device and any age group. Patient triage, the second scenario After the initial pandemic peak, we built a second Healthbot scenario for patient triage. As hospitals began receiving non-COVID patients again, they needed a tool to risk-profile each patient before their appointment. We built a chatbot-driven survey, delivered via SMS, with results linked directly to the patient's Electronic Patient Dossier (EPD). The first deployment was at Ziekenhuis Oost-Limburg (ZOL). IMPLEMENTATION #### Six hospitals live in twelve days Between March 25 and April 6, 2020, we deployed chatbots at six Belgian hospitals: AZ Rivierenland, Ziekenhuis Waregem, SLBO, AZ Delta, AZ Vesalius, and ASZ Aalst. Each chatbot guided users through a symptom self-assessment, then recommended a course of action based on their responses. After the assessment, users were asked for basic demographic data — age, sex, postal code. This data fed into Power BI dashboards that gave hospital management a view on potential admission volumes. Arinti's fast approach and smooth communication ensured that the COVID-19 symptom checker on ASZ.be was accessible to the public in no time. The chatbot not only provides answers to questions related to possible symptoms, but also tells the visitor what to do if they show certain symptoms. 6 Hospitals deployed between March 25 and April 6, 2020 700+ Self-assessments completed within 4 days at one hospital 12 days From first deployment to sixth hospital live #### Impact We are very enthusiastic about how we were able to build a very useful application in a very short term to support our covid triage flows. In a few days and with a minimum of costs, we had an accessible and patient-friendly application operational. --- ### [Social robots for personalised diabetes education in children](https://arinti.ai/cases/robocure) Client: ROBOCURE consortium · Published 2020-01-01 > ICON research project investigating how social robots, IoT, and AI can improve diabetes treatment in children. We built StoryLine 2 Pepper: a tool that lets non-technical staff create interactive learning content delivered by a humanoid robot, with LRS integration for monitoring patient education. CONTEXT #### What ROBOCURE set out to do ROBOCURE investigates how social robots can inform and guide patients, using patient data collected via the robot to establish more personalised treatments. The project studies how machine learning and medical expertise can automate data management and analysis for personalised therapies. The research focuses on diabetes treatment in children. The project is part of the ICON applied research programme and includes partners from both the private and public sector: Arinti, The Learning Hub, QBMT, UZ Brussel, Medtronic Belgium, imec — IDLab — UGent, imec — SMIT — VUB, VUB — GRON, and VUB — R&MM. ROBOCURE researches and validates the use of robots, the Internet of Things and artificial intelligence in healthcare to improve medical therapies. The project is based on remotely collected data. Why a robot Research at Plymouth has shown that children respond well to robots when it comes to influencing behaviour and opinions, making them suited for the role of educator. A humanoid robot combines the benefits of digital content with reactive, social interaction based on the child’s behaviour. The solution also supports browser access when no robot is available. How it works We built on StoryLine, a widely used e-learning authoring tool similar to PowerPoint. Learning content created in StoryLine is exported to HTML with integrated LRS (Learning Record Store) support. We added an interface layer that enables interaction with the Pepper robot through just two commands: Listen (for speech recognition) and Say (for robot speech). Browser support uses the Web Speech API and SpeechSynthesis for testing without a physical robot. Patient interaction When a learning module starts, Pepper introduces itself and asks the patient to proceed. The tablet hints at expected words to improve recognition. If recognition fails, the robot asks the patient to repeat. Tablet interaction serves as a fallback. Modules are deployed to the robot via a task scheduler built by imec, with LRS monitoring built in. IMPLEMENTATION #### StoryLine 2 Pepper: e-learning content for robots We partnered with The Learning Hub to merge e-learning with social robotics. The goal: allow non-technical staff to create interactive learning content that a humanoid robot could deliver, without specialist programming. RESULTS #### What we delivered The project produced a pipeline for creating learning content without technical expertise, delivered by a humanoid robot. StoryLine’s interactive features augment standard e-learning functionality. Educational results are monitored through native StoryLine and LRS integration. At the time of the project, the integration supported the Pepper robot, with the architecture designed to extend to other robots that have a tablet and JavaScript API. ROBOCURE received coverage from VRT NWS, VTM Nieuws, HLN, Nieuwsblad, and RTBF. --- ### [Chatbot Bertje - digital citizen services for the City of Roeselare](https://arinti.ai/cases/chatbot-bertje) Client: Stad Roeselare · Published 2019-10-01 > Chatbot Bertje answers citizen questions for the City of Roeselare across all municipal domains, 24/7 — built with a database of 1,000 questions, NLP-based training, and automatic routing to the city's complaint handling system. Live since October 2019. What the city wanted The goals were concrete: a digital service channel available 24/7, a reduction in phone calls to the central city number (1788), a new contact channel alongside phone, email, and social media, and — eventually — a replacement for the website’s search function. CONTEXT #### A city investing in digital-first services Roeselare had made citizen services a strategic priority since 2013 and adopted an omnichannel approach: encouraging citizens to start online, then call if needed, and visit in person as a last resort. In 2017, the city worked with iMinds (imec) on a Smart City assessment, which identified a chatbot as a viable pilot project. From 2017 to 2020, Roeselare participated in Like, an Interreg North Sea Region programme focused on digital innovation in public services. A chatbot fitted directly into that programme's scope. How Bertje works Bertje uses a knowledge base of roughly 1,000 questions and synonym variations. The chatbot is continuously trained through natural language processing to recognise questions and provide accurate answers. Staff can control the confidence threshold: only answers scoring above a set level are shown to citizens. The system logs every interaction, and staff train the chatbot by reviewing and scoring its responses. Routing and workflows When a citizen wants to report an issue, Bertje routes them directly to the city’s complaint handling system. For specific question types, custom conversation flows guide the citizen through the relevant process. IMPLEMENTATION #### Building and training Bertje We built Bertje in partnership with Cronos, with Microsoft providing Azure infrastructure support. The project was co-financed through the Interreg NSR programme (50%) and involved investment from all partners. The total cost for the city was approximately €22,000. Staff investment The project required approximately 0.8 FTE over one year from city staff for content development and chatbot training. That investment also resulted in a thorough review and update of the city’s website content. RESULTS #### What Bertje delivers Bertje has been live on roeselare.be since 1 October 2019. The chatbot covers the city’s full range of municipal services rather than focusing on a single topic. Citizens get answers around the clock. The chatbot also gives the city better insight into what citizens actually ask, which helps train the contact centre staff. And because Bertje pulls information from the website, it forces the city’s teams to keep their web content up to date — an unintended but valuable side effect. --- ### [VDAB Future of Work Challenge](https://arinti.ai/cases/vdab-future-of-work-challenge) Client: VDAB · Published 2019-06-01 > Winner of the VDAB Future of Work Challenge with CompeTrend, a system that predicts competency trends and maps skill similarity to help jobseekers find realistic retraining paths. CONTEXT #### A labour market under pressure The Flemish labour market had a low unemployment rate (6% in March 2019) and a rising activity rate (72% in December 2018), but filling vacancies was becoming harder. The skills jobseekers had often did not match what employers were looking for, and new technologies — automation and AI — were expected to widen that gap further. VDAB had defined four strategies to address this: skills upgrading, retraining, activation, and productivity increase. The Future of Work Challenge was designed to bring in outside perspectives — 33 organisations submitted proposals, and VDAB selected four winners. It is important for us to work with an open mind. The Flemish labor market will face new challenges in the next 10 years. We suspected that with all the data we have available, innovative solutions are possible that we hadn't thought of ourselves. Trend analysis We analysed the evolution of competencies in both vacancies and CVs. The approach involved clearing seasonal cycles and random deviations from the data, then predicting short- and long-term evolution using time series methods (VARIMA and recurrent neural networks). Competencies were classified into categories — disappearing, declining, stable, growing — to surface the trends that mattered for retraining policy. We also mapped the mismatch between supply and demand: the ratio of vacancies requesting a given competency to CVs listing it. Skill similarity The second component mapped data-driven similarities between competencies using association rule mining on CV data. If two competencies frequently appeared together in CVs, the system scored them as similar. For example: Python and R appeared together with 70% probability, yielding a skill similarity score of 0.8. These similarity scores fed into hierarchical clustering, grouping competencies that the existing Competent 2.0 taxonomy had not yet connected — giving VDAB a data-driven basis for improving the taxonomy itself. IMPLEMENTATION #### What we built: CompeTrend CompeTrend worked with Competent 2.0, the standard competency framework used by VDAB. The system had two components: trend analysis and skill similarity. RESULTS #### What CompeTrend delivered The project supported two of VDAB's four core strategies: skills upgrading and retraining. Concretely, the system provided trend predictions per competency, allowing VDAB to identify which skills were declining before they disappeared entirely. Similarity mappings between competencies showed jobseekers which skills they already had that could serve as stepping stones to future-oriented competencies. The clustering output also gave VDAB a data-driven basis for improving the Competent 2.0 taxonomy, and enabled faster orientation of jobseekers toward feasible and relevant training paths. --- ### [Business process automation for HR services](https://arinti.ai/cases/partena-professional) Client: Partena Professional · Published 2019-02-01 > Chatbot Louise automates end-of-contract procedures at Partena Professional — reducing dismissal advice from 60+ minutes to 15–20 minutes, serving 900 payroll consultants daily, and going from prototype to production in three months. CONTEXT #### A manual process under regulatory pressure The Legal Partners team at Partena Professional assists customers with end-of-contract procedures: dismissal letters, mutual agreements, contract terminations. For each case, an HR consultant had to verify compliance with dismissal law (outplacement, early retirement, industry-specific rules), calculate the notice period and resignation fee, register the case in Topdesk, and manually draft the resignation letter. The data was typically captured on paper during a phone call, then entered into multiple systems. The full process took 60 minutes or more, depending on complexity and consultant experience. Regulatory deadlines added pressure — if a dismissal needed to start on a Monday, the letter had to leave on Wednesday. What Louise does Louise captures the client's case data through a conversational interface, then automates the downstream steps: calculating the dismissal period, creating the Topdesk ticket, and generating a custom resignation letter. The letter is sent to a Legal Consultant for final validation before going to the customer. Fast adoption After one month of intensive testing, Partena Professional decided to put Louise into production. By February 2019, Louise was live and integrated into the Legal Partners team's daily operations. IMPLEMENTATION #### Louise: from prototype to production in three months We started building the prototype of Louise in the autumn of 2018. After a few weeks of development, Partena's payroll consultants began testing the chatbot with real client data. By investing in a scoped project that was used throughout the company, the adoption rate of new technologies grew enormously. Thanks to the flexibility and effort of Arinti, we were able to prove the value of chatbots in a few weeks. RESULTS #### What changed Dismissal advice that previously took 60+ minutes now takes 15–20 minutes of consultant review — roughly three times faster. The error rate dropped significantly. HR consultants report that offloading the repetitive parts of the process frees them for what they were hired to do: person-to-person interactions with customers. Louise is used daily by roughly 900 payroll consultants and is considered a colleague of the Legal Department. In 2020, we developed a public-facing version called Alex, available on Partena Professional's website, allowing customers to generate dismissal letters independently. ## Insights ### [Three frontends rebuilt in three weeks, with AI agents doing the heavy lifting](https://arinti.ai/insights/three-frontends-rebuilt-in-three-weeks-with-ai-agents) Category: article · Published 2026-08-06 > An external pentest turned a long-postponed frontend rebuild into a hard deadline. One developer, working with AI agents, delivered three modern, secure and mobile-friendly web apps in three weeks, ten days early. CASE STORY #### From pentest to three renewed frontend applications For a Belgian financial services provider in the automotive sector, the frontend rebuild of their dealer performance platform had been on the wish list for a while. The three web applications ran on Vue 2, a framework out of long-term support since the end of 2023, meaning no more security updates. Then an external penetration test flagged exactly that. Not a breach, not an incident: an unsupported framework, written up as a security finding. The wish became a project with a hard deadline, off the unsupported stack before the summer holidays. Three weeks of work later, all three applications were rebuilt: modern, secure and mobile-friendly, running against the live backend ten days before that deadline. Not by a project team, but by one developer working with AI agents. This is what AI-assisted development looks like when you take quality as seriously as speed. Here is what it delivered for the client, and how. THE STARTING POINT #### A business-critical platform on an unsupported framework The platform is business-critical: dealers and salespeople use it daily to follow up objectives, commissions, contracts and invoices, and the client's admin team manages the entire commission engine behind it. The backend had been migrated to .NET 10 earlier in the year, but the three frontends still ran on Vue 2: end of life since 2023, meaning no more security fixes, whatever a scan turns up next. An external penetration test made that concrete. Nothing had gone wrong, but the report flagged the outdated frontend components as a security finding, with a clear recommendation: get off the unsupported stack. The refactoring wish that had been quietly postponed suddenly had a hard deadline: finished before the summer holidays. This is the kind of project organisations dread, and postpone, for a reason. The scope covered a management cockpit full of KPI dashboards, commission views and reports, a mobile app for salespeople, and an admin application with over 30 screens. The classic answer to a rebuild of that size is a team working for six months or more. The client got a different answer. Security patching that actually works Vue 2 no longer receives security fixes, so patching the old stack was painting over a crack. The new frontend runs on current, maintained frameworks, and because the three applications share one repository and one set of dependencies, a vulnerable package is updated once and every app is fixed. The structured patch management the pentest recommended is now realistic instead of theoretical. Mobile friendly by default Every screen was rebuilt on responsive components, so dealers and salespeople can now check their objectives and commissions on a phone as comfortably as on a desktop. For the people in the field this was one of the most tangible wins of the project, and it came as part of the same rebuild rather than as a separate project to budget for. Dark mode came along in the same move. Changes land faster, in all three apps at once The three applications now share one component library: tables, filters, cards and forms are built once and reused everywhere. A new screen is largely assembled from existing parts, and an improvement to a shared component lands in all three apps at the same time, on this project and on every change after it. The demo week proved the point: over 30 change requests went live within days. A smaller maintenance bill every year after The legacy applications lived in separate codebases and carried years of accumulated packages, many no longer maintained. The rebuild started from a clean slate: one repository (with the old git history preserved), a deliberately small set of current libraries, and all three apps permanently on the same versions. An upgrade now happens once and is tested once, instead of three times: a smaller attack surface today, and a smaller upgrade bill every year. Documented for whoever comes next Every component has a written design spec and implementation plan in the repository. The next change starts from documentation instead of code archaeology, whether a human developer or an AI agent picks it up. THE RESULT #### What the rebuild delivered The deadline was met, and the pentest finding that started it all is closed. But the three weeks delivered more than a like-for-like copy on a newer framework. THE ORCHESTRATION #### One developer, a fleet of agents The agents produced the volume. The specs, gates and reviews kept it correct. THE APPROACH #### Speed from agents, quality from structure Speed came from AI agents. Quality came from the structure we put around them. We started with an inventory: every screen and component of the three legacy applications was catalogued, 46 components in total, and each one became a separate ticket. For every ticket, an AI agent first wrote a design spec, then an implementation plan, and only then the code. Every commit was reviewed against its task, and the branch as a whole got a final security review before merging. The same flow had proven itself a few weeks earlier, when we used it to remediate the penetration test findings in the backend. The architecture did the rest. Instead of three separate codebases, the new frontend is one repository with a shared component library: a table, a filter or a form is built once and used by the cockpit, the salesperson app and the admin application alike. All screens were first built as clickable pages against mock data (in Storybook), so layout and behaviour could be checked early, and only then connected to the live .NET backend. Automated checks (type checks, linting and build verification) ran on every change, and a human developer reviewed and live-tested everything against the real system. That combination let one developer orchestrate what is normally a team effort. The agents produced the volume; the specs, gates and reviews kept the volume correct. 9 July The shared foundation (navigation, filters, tables) and the full set of cockpit screens were built and connected, a day before the official start. 10 July All six salesperson screens were finished and the entire admin application existed as clickable screens in Storybook. 20 July All admin screens were connected to the live backend. Every screen of the three applications was built and working, ten days before the deadline. 27 July Demo with the client. 30 July Deadline. More than 30 feedback items from the demo rounds had been picked up and resolved, most within days of being raised. Because every component follows the same spec and plan structure, feedback lands in small, well-defined tickets that an agent can pick up immediately. The client saw the remarks from Monday's demo live on the test environment that same week. THE TIMELINE #### Three weeks, measured The project was planned from 10 to 30 July. The actual pace: 1 Developer, orchestrating a fleet of AI agents 3 Business-critical web applications rebuilt ~3 weeks From kick-off to delivered platform 10 days All screens built and connected ahead of the deadline 30+ Demo feedback items resolved within days IMPACT #### The project in numbers THE QUALITY QUESTION #### Quality is the honest concern, so we engineered for it The honest concern with AI-generated code is quality, and it is a concern we share. That is exactly why every ticket gets a spec before a single line is written, why every commit is reviewed, and why the final branch went through a dedicated security review. This is the first AI-built project that doesn't look AI-built. YOUR LEGACY FRONTEND #### What this means for your legacy frontend Frameworks age faster than the applications built on them. If your frontend runs on Vue 2, AngularJS or another end-of-life stack (a framework that no longer receives security updates), you are carrying a growing security risk and maintenance bill. And the classic answer, a rebuild project of six months or more, is exactly why most companies keep postponing it. Our experience with this project says the trade-off has changed. With AI agents inside a disciplined spec, review and verification process, a rebuild of this size fits in weeks, not months, without cutting corners on quality or security. If you have a modernisation that keeps sliding down the backlog, talk to us. We will gladly show you how we work, demo included. --- ### [Advancing Your Data Governance Strategy with Microsoft Fabric](https://arinti.ai/insights/microsoft-fabric-data-governance) Category: article · Published 2024-09-15 > Data governance is no longer just about securing information or meeting compliance standards. With Microsoft Fabric, organisations can bring together fragmented tools into one cohesive governance framework. Exploring data governance through Fabric Data governance involves organising and securing data so it can be trusted, compliant, and readily accessible. With Microsoft Fabric, this means bringing together previously fragmented tools and processes into one cohesive framework. At its heart, data governance ensures the right data is in the right hands at the right time, while maintaining integrity and security. Overcoming fragmentation Fabric helps organisations handle fragmentation across scattered systems by consolidating data sources, catalogs, and governance policies into a unified platform. This eliminates the silos that typically hinder data discovery and compliance. Meeting compliance requirements With evolving regulatory requirements, organisations need governance frameworks that adapt quickly. Fabric's built-in compliance tools help teams stay ahead of regulatory changes while maintaining consistent data handling practices. Establishing data trust Building confidence in data accuracy and integrity is essential. Fabric provides the tooling needed to verify data origins, track transformations, and ensure that stakeholders can rely on the information they access. DATA GOVERNANCE Data governance is no longer just about securing information or meeting compliance standards. It's about ensuring that data remains a trusted, transparent, and actionable resource. #### Key features that support governance Fabric integrates multiple capabilities for data governance. Data lineage tools help users trace the origin, movement, and transformation of data — vital for audits, troubleshooting, and compliance. Integration with Microsoft Purview enables tagging, tracking, and managing data through sensitivity labels and compliance monitoring. AI-driven discovery makes it easier to locate relevant data in large datasets where manual searches are impractical. Provenance and auditability Provenance tracking verifies data origins and transformations, creating a clear audit trail. Combined with built-in auditability features, organisations can demonstrate compliance and accountability at every stage of the data pipeline. Integrity and safeguards Integrity checks safeguard against corruption or mismanagement, ensuring that data quality is maintained as it flows through different systems and processes within the organisation. AI-enhanced governance workflows Generative AI introduces new possibilities for automating and enhancing governance workflows within Fabric — from repetitive task automation to smart insights for faster identification of patterns and anomalies. This allows governance teams to focus on strategic decisions rather than manual oversight. #### Trust and Generative AI in governance Trust is central to any governance strategy. Fabric incorporates a range of mechanisms to ensure data remains reliable and accountable throughout its lifecycle. From traditional to modern frameworks By incorporating AI, transparency, and scalability into traditional frameworks, organisations can build governance strategies that grow with their data needs rather than becoming obstacles to innovation. Data as a strategic advantage This shift allows organisations to use their data as a strategic advantage while maintaining stakeholder confidence. When governance is embedded into the data platform itself, it becomes an enabler rather than a bottleneck. #### Understanding governance's role Data governance is not about tools alone — it's about the processes, principles, and practices that keep data aligned with organisational goals. Platforms like Microsoft Fabric show how governance can evolve to meet modern needs. #### Did you know that of all information within companies, an estimated 80% is unstructured? Modern governance platforms like Microsoft Fabric help organisations bring order to this chaos — turning unstructured data into a trusted, actionable asset. --- ### [Redefining data management for organisations with Microsoft Fabric](https://arinti.ai/insights/microsoft-fabric-data-management) Category: article · Published 2024-08-01 > Microsoft Fabric unifies diverse data management tools into a cohesive SaaS platform. This article explores its architecture and how they empower organisations to optimise data operations. Structuring for success with domain orientation A Fabric domain is a logical grouping of workspaces. They present a convenient way to manage access and permissions to groups of resources — for example, a domain for Sales and another for Marketing. Subdomains further isolate management concerns, allowing tailored oversight for specific functions. The domain ownership principle mandates domain teams to take responsibility for their data, moving analytical and operational data ownership away from the central data team. Collaborative workspaces The collaborative workspaces within Microsoft Fabric provide a unified environment for data ingestion, machine learning, real-time analytics, and reporting. Managed by regions and capacities, they offer precise control over resource allocation and compliance with regional data regulations. The role-based access control system — encompassing admin, member, contributor, and viewer roles — ensures data access is secure and appropriately managed. MICROSOFT FABRIC Fabric's SaaS architecture simplifies infrastructure complexities. Activated via the Power BI Admin Portal, the platform ensures swift deployment, scalability, and automatic updates — allowing organisations to focus on actionable insights rather than operational details. #### OneLake: unified storage at the core OneLake, the backbone of Microsoft Fabric, unifies data management with a centralised, scalable framework built on Azure Data Lake Storage Gen2. By integrating workspaces, lakehouses, and warehouses, it eliminates silos and simplifies governance, security, and compliance. With lightweight data virtualisation, OneLake enables real-time access without duplication, enhancing operational efficiency across teams and domains. #### Security model The security model in Microsoft Fabric is multi-layered: platform-level authentication via Microsoft Entra ID with MFA and conditional access, domain-specific configurations for tailored security settings, workspace-level controls for resource management, and item-level access for granular data protection. --- ### [From scattered onboarding emails to a single chatbot](https://arinti.ai/insights/building-internal-hr-chatbot-genai) Category: guide · Published 2024-03-15 > We built an internal chatbot to answer employee questions using OpenAI GPT, LangChain, and our own Notion knowledge base. Here's how we did it, from document ingestion to deployment. How it works — three steps First, Document Ingestion: we convert all Notion onboarding content into numerical vectors using OpenAI embedding models. The content is split into chunks using LangChain's markdown splitter, then stored in a vector database. Second, Query: the user's question is converted into a vector using the same embedding model, then matched against the vector database via similarity search. The most relevant content chunks are passed along with the question to OpenAI GPT 3.5 Turbo, which formulates an answer based on a structured prompt. Third, Memory: the chatbot tracks conversation history, combining previous messages with new questions into standalone queries. GEN AI When new colleagues join Arinti, onboarding information used to be scattered across emails and documents. With Large Language Models now available, we built a chatbot that answers employee questions instantly — powered by our own Notion knowledge base. #### Automation and deployment Azure Functions automatically fetch new Notion content on a daily basis, keeping the chatbot's knowledge up to date. The front-end is built with Streamlit and embedded directly into our Notion workspace — so employees access it without switching tools. The same architecture works for any knowledge base: FAQ pages, project documentation, or policy documents. As long as the content is stored somewhere, it can serve as the foundation for a chatbot. --- ### [Interview with Microsoft on it's three-pillar sustainability strategy](https://arinti.ai/insights/interview-microsoft-sustainability-partner) Category: article · Published 2023-03-01 > In the third instalment of our sustainability interview series, we spoke with Myriam Broeders (CTO) and Mathias Vergauwen (Technical Architect) at Microsoft Belgium about their three-pillar sustainability strategy. Sustainable by design Microsoft's new Belgian Azure region data centres consume only 12 percent on top of the actual server's power draw — a radical efficiency improvement compared to traditional data centres. The buildings use modular steel construction for long-life span with minor alterations, and servers are repurposed, refurbished, or properly recycled rather than shredded. On the software side, Microsoft is heavily involved in the Green Software Foundation, pushing towards sustainable software development. SUSTAINABILITY SERIES Microsoft introduced an internal carbon tax in 2009 and pledged to reverse all historic carbon emissions by 2050. We spoke with Myriam Broeders, CTO and responsible for sustainability strategy at Microsoft Belgium & Luxembourg, and Mathias Vergauwen, Technical Architect in the Microsoft Technology Center in Brussels. Microsoft's sustainability policy rests on three pillars: Sustainable by design, Sustainable platform, and Sustainable collaborations. #### Cloud for Sustainability Microsoft's Sustainability Manager SaaS product enables businesses to collect emission data in one place, map it across countries and departments, and generate reports and dashboards. Currently focused on greenhouse gas emissions, it will soon incorporate water use and waste management data. As Myriam notes: getting started is often the hardest part. Most sustainability journeys start with Excel files and high-level benchmarks. Even partial data is a valid starting point. Sustainable collaborations Microsoft runs the AI For Good project and the Entrepreneurship for Positive Impact program for start-ups with positive societal impact. Belgian examples include BeeoDiversity (biomonitoring with bees) and Ayes (AI-powered accessibility for blind people). Microsoft's internal carbon price stands at 15 dollars per ton CO2, with higher rates for easily avoidable emissions like air travel (over 100 dollars per ton). --- ### [Interview with BekaertDeslee on circularity challenges and the role of digital tools](https://arinti.ai/insights/interview-bekaertdeslee-sustainability-technology) Category: article · Published 2023-01-15 > In the second instalment of our sustainability series, we spoke with Charles Christiaens, sustainability manager at BekaertDeslee, about circularity challenges in textile manufacturing, CSRD preparation, and the role of digital tools. Circularity as the biggest lever Mattress fabrics are composed of hard-to-separate materials — composite textiles with cotton, polyester, and elastane, plus moisture-rejecting coatings. This makes recycling extremely difficult today. But BekaertDeslee is working with recycling companies and researchers to prepare products for end-of-life recycling, anticipating that technology will evolve significantly within the 8-10 year mattress lifespan. SUSTAINABILITY SERIES BekaertDeslee, founded in 1892, is a textile manufacturing company specialising in mattress fabrics. They are the global leader in their field, with production facilities in all major markets. We spoke with sustainability manager Charles Christiaens about their challenges, ambitions, and preparation for new EU regulations. The textile industry is the fourth most polluting industry in the world, and mattresses are landfilled over 90% of the time. As Charles explains: making a sustainability shift in a market that isn't ready for these changes is, well, unsustainable. #### CSRD: a serious challenge BekaertDeslee did a dry run for the upcoming European CSRD. Charles's key takeaway: companies need to take these reports as seriously as financial reporting. An auditor will challenge every evolution in your numbers, and you need data systems with feedback loops — manual input doesn't cut it. Still, Charles believes the regulations are necessary: if nobody's watching, nothing will change. Where digital innovation helps Data capturing and measuring impact remains challenging. BekaertDeslee has a well-integrated SAP that helps, but some data still needs manual processing. AI-driven software can help with predictive reporting, filtering and merging data, and bringing the cost of sustainability reporting down. As Charles puts it: new technology will need to enable automating more of these processes. --- ### [Interview with Dyneema - Growing towards a sustainable, circular industry model](https://arinti.ai/insights/interview-dyneema-avient-circular-industry) Category: article · Published 2022-11-01 > In the first instalment of our sustainability series, we spoke with Keith Rowsey, sustainability manager at Avient Protective Materials (Dyneema), about building circular value chains, reverse logistics, and scope 3 emissions. Reverse logistics and recycling For every market segment — rope, ballistics, consumer protection — Avient works on two major steps: getting products back and finding the technical solution to recycle them. In the rope industry, they're discussing collection points at major ports. But products are often processed, dyed, and combined with other materials, so the recycling solutions need to accommodate this. As Keith notes: the shorter the recycling loop, the better. SUSTAINABILITY SERIES Avient produces extremely strong Ultra High Molecular Weight polyethylene (UHMWPE) fibres, branded as Dyneema. Used in industrial roping, ballistic protection, and consumer products. We spoke with Keith Rowsey, sustainability manager, about the challenges and intricacies of sustainability in manufacturing. Avient set ambitious 2030 goals: all Dyneema products made of at least 60% bio-based or recycled feedstock. Their production sites in Heerlen (NL) and Greenville (NC, USA) already run on 100% purchased renewable electricity. They also helped launch CirculariTeam, an industry coalition for building circular value chains. #### Data and traceability Longer use and repurposing are where the biggest gains lie. Avient designed DM20, a fibre with less creep that lasts 25+ years for offshore wind farms. But enabling circular business models requires extensive data and traceability: how many lifts has a rope been used for? When was it produced? What materials are inside? RFID-chipping products for end-of-life identification is one of the approaches being explored. Scope 3 emissions progress For industrial companies, roughly 75-80% of total emissions are scope 3. Three years ago, if you asked someone about their carbon footprint, they often didn't know or were unwilling to share. Now it's a given for doing business. Still, gathering and aggregating all that information into a correct calculation remains a major effort. Keith's biggest wish: standardised terminology so everyone understands what kind of data they're actually dealing with. --- ### [Getting value from large datasets in a short amount of time](https://arinti.ai/insights/getting-value-large-datasets-short-time) Category: guide · Published 2022-06-01 > We performed a one-week data analysis for NMBS (Belgian railways) on datasets with millions of rows. Here are our tips for investigating large datasets when time is of the essence. Set up your workflow correctly from the start Before exploring data, ask: how much data do I have? How will I present results? Does the analysis need to be reproducible? For NMBS, we set up a pipeline on day one: data from Azure Blob Storage to Databricks, output back to Blob Storage, with Azure Data Factory triggering on new files. A Power BI dashboard sat on top for both exploration and final presentation. This ensured reproducibility, boosted productivity, and let the client run future analyses by simply uploading new files. DATA ANALYSIS One week, millions of rows, six key questions. We were asked to perform a short data analysis for NMBS, the Belgian railway company. Over one week, we investigated two related datasets (several million rows each) to help them understand their data and identify missing observations. The analysis had to be reproducible and presentable by the client afterward. Here are our tips for investigating large datasets when time is limited. #### Focus on specific examples When faced with large volumes, focus on specific examples rather than just summaries. Simply saying '15% of data is missing' isn't very helpful. By investigating examples of both problematic and non-problematic observations and discussing them with the business, we discovered that many initially flagged cases were 'missing but non-problematic' — which were then excluded from the final analysis. Communication, specific goals, and documentation With limited time, daily calls and close contact with the client prevent wasted hours on low-priority work. We agreed on six specific questions upfront and presented conclusions at the end of the week. We spent our last half-day on documentation alone — ten percent of total worktime, but without it the entire analysis would have been practically worthless for future use. --- ### [Ethics and the power of AI – having a chat with Minister Bart Somers](https://arinti.ai/insights/ethics-ai-minister-bart-somers) Category: article · Published 2022-03-01 > The mayor of Mechelen and Vice-Minister-President of the Flemish Government met with Arinti and Microsoft to discuss the possibilities of AI in government and its ethical consequences. The power of AI in government AI saves time and money by automating repetitive tasks. But as Sammy Deprez (Arinti) explains, it can also reduce stress, provide new insights, and prevent frustration. The city of Mechelen's chatbot Yana is an example: it checks in with young people going out at night, simulates phone calls, and puts them at ease. Pre-built AI blocks make it possible for organisations to get started right away. Arinti helps optimise these blocks and adapt them to specific needs. AI ETHICS On 10 February 2022, Minister Bart Somers — mayor of Mechelen and Vice-Minister-President of the Flemish Government — met with Arinti and Microsoft to discuss AI possibilities and ethical consequences. Somers envisions a future where citizens can handle all government services from anywhere, without physically visiting a town hall. The Government of Flanders launched a project to make local administrations digital-proof. But as Somers clarifies: physical service should always remain possible. By doing more digitally, capacity is freed up for personal, customised service provision. #### Responsible AI and ethical frameworks People quickly think in doomsday scenarios with AI, but computers are still just computers. To deal with concerns about privacy, control, and transparency, organisations need ethical frameworks. These already exist at European and global level, where cities and countries share knowledge and experiences. With deep learning, you can go so deep that you no longer know how something was calculated. As Koen Everaert (Arinti) warns: if you make predictions based on historical data, you have to make sure the data is right. Microsoft's framework for responsible AI is built on six principles: fairness, reliability, privacy, inclusivity, transparency, and accountability. --- ### [Internship report: Using AI to personalise medical questionnaires](https://arinti.ai/insights/internship-ai-personalize-medical-questionnaires) Category: article · Published 2021-09-01 > Intern Usman Dankoly researched AI techniques to personalize PRO-CTCAE questionnaires for lung cancer patients, using K-means clustering and personalised recommendation scores to reduce questionnaire burden. Approach: clustering and recommendation Working with a dataset of 35 lung cancer patients (544 rows × 55 columns), intern Usman Dankoly applied K-means clustering on tumor classification and WHO scores, identifying two distinct patient clusters. The importance of each question was calculated by frequency of non-zero composite scores, while uncertainty was measured via standard deviation within clusters and across time. INTERNSHIP Personalising cancer patient questionnaires with AI. PRO-CTCAE is a patient-reported outcome system for evaluating symptomatic toxicity in cancer clinical trials. The library includes 124 items representing 78 symptomatic toxicities. Patients fill these in weekly during key treatment periods. The research objective: instead of a one-size-fits-all selection of questions, use AI to recommend the 10 most relevant questions per patient — those that matter most and are least known. #### Personalised results Individual recommendation scores combined importance and uncertainty as tunable hyperparameters. In early treatment phases, more weight goes to uncertainty (what we know least). In later phases, more weight goes to importance (what matters most for this specific patient). This reduces questionnaire burden while maintaining clinical relevance. — Usman Dankoly, EHB Brussel --- ### [Internship report: Solving a game of Qwirkle with Custom Vision & OpenCV](https://arinti.ai/insights/internship-qwirkle-custom-vision-opencv) Category: article · Published 2021-07-02 > An internship project using Custom Vision and OpenCV to detect Qwirkle tiles and calculate optimal moves. Colour and shape detection Rather than training Custom Vision to identify colour-shape combinations separately, the team used pixel analysis at bounding box centres, converting RGB values to HSV colour space for accurate colour determination independent of lighting conditions. Custom Vision was trained on six shape categories using approximately 53 photos per classification, achieving reliable performance through variable-distance training samples. INTERNSHIP REPORT This 50-day internship project investigated using artificial intelligence to determine optimal moves in the board game Qwirkle, combining computer vision technologies with game logic algorithms. This contribution comes from Evi Leroy at VIVES Hogeschool. #### Perspective correction and preprocessing OpenCV handled perspective correction before Custom Vision analysis. The preprocessing pipeline included edge detection, contour analysis, and Hough line detection to orient game boards correctly for recognition. Brute-force optimisation for maximum points A brute-force algorithm evaluated all possible tile placements, scoring combinations based on Qwirkle's point system and identifying sequences yielding maximum points. The three-phase methodology — research, training, and implementation — demonstrated how Custom Vision API and OpenCV can work together for practical game-solving applications. --- ### [Cognitive Search: unlocking value from unstructured data](https://arinti.ai/insights/cognitive-search-unstructured-data) Category: article · Published 2021-07-02 > How Tafuta, Arinti's cognitive search platform, helps organisations retrieve value from unstructured data using NLP, speech-to-text, and image processing. AI-powered search beyond keywords Unlike traditional search engines that rely on keywords and metadata, Tafuta incorporates Natural Language Processing, speech-to-text conversion, and image processing to understand data context and improve result relevance. It consolidates scattered, unstructured data — including scanned PDFs, images, and audio files — into one searchable location. COGNITIVE SEARCH Approximately eighty percent of company information exists in unstructured formats. Tafuta — Swahili for 'search' — is Arinti's cognitive search platform designed to help organisations retrieve all this data in an intuitive, efficient way. #### Ingestion, indexing, querying The system operates through three stages. During ingestion, databases absorb new data. During indexing, information enters the search solution with entity recognition and metadata creation. During querying, users retrieve connected results. Customisable skills allow organisations to tailor solutions for specific industries or domains. From police investigations to trade intelligence In December 2020, Arinti won the Belgian Federal Police's Smart Policing Hackathon by developing investigation techniques using Tafuta's Graph Analytics feature. The solution converted non-searchable content like phone recordings and camera images into searchable text, helping investigators visualise criminal connections. Organisations including Flanders Investment & Trade, VITO, VLAIO, and VIB have adopted the platform. --- ### [Internship report: Education Recommender Bot](https://arinti.ai/insights/internship-education-recommender-bot) Category: article · Published 2021-07-01 > Intern Brecht Noyens built a chatbot that recommends the best field of study to future students, combining Microsoft Bot Framework Composer with a Python-based recommendation engine on Azure. Architecture The chatbot was built with Microsoft Bot Framework Composer and published on Azure. Education data from onderwijskiezer.be was scored across aspects like science, mathematics, politics, and statistics. The recommendation engine runs as a Python Azure Function, connected to an Azure database. The bot collects user preferences through conversation, then calls the recommendation API to suggest matching fields of study. INTERNSHIP As part of his internship for Applied Computer Science (Specialisation AI) at Thomas More Hogeschool Geel, Brecht Noyens developed a chatbot that helps future students select the most appropriate field of study based on their interests. #### Extensions The bot supports both Dutch and English. A QnA Maker service was integrated so users can ask general questions during the conversation. A speech-to-text feature was added using Azure Speech Services. The final product combines adaptive questioning, a solid recommendation engine, and multilingual support in a single conversational interface. — Brecht Noyens, Thomas More Geel --- ### [Redeeming the chatbot: more than easy customer interactions](https://arinti.ai/insights/redeeming-the-chatbot) Category: article · Published 2021-06-01 > Chatbots are more than marketing gimmicks. From virtual assistants and safety buddies to symptom checkers during COVID-19: here's an overview of different chatbot types and how we build them. The chatbot as virtual assistant Every organisation has straightforward tasks that take too much time. As a digital employee, a chatbot helps human colleagues by executing tasks or parts of tasks. It continuously improves through new conversations — and it never forgets. CONVERSATIONAL AI With machine learning and NLP technology, chatbots can do far more than answer a few simple questions. Before development starts, we follow a structured trajectory: defining goals, determining how smart the chatbot needs to be, then running a Bot Insights Workshop to determine intents and entities. Only after prototyping do we move to development. Customers expect more from chatbots today. And the technology delivers. #### The chatbot as buddy Chatbots don't have to serve a marketing purpose. YANA (You Are Never Alone), built for the city of Mechelen, gives youngsters and their parents a sense of security when going out. Through Facebook Messenger, YANA tracks travel plans via Google Maps integration and notifies contacts when you've arrived safely. The chatbot as information source Bertje, built for the city of Roeselare, answers over a thousand questions. Using NLP, it recognises unknown questions and flags them for future learning. During COVID-19, we were selected as Microsoft's exclusive Belgian implementation partner for the HealthBot. In ten days, six healthbots went live in Belgian hospitals, integrated with Electronic Patient Dossiers for risk profiling. --- ### [How to: Installing modules in a Python Azure function](https://arinti.ai/insights/installing-modules-python-azure-function) Category: guide · Published 2021-04-01 > A practical guide to installing Python modules in Azure Functions, including setup tips and deployment considerations. Setting up your environment Initial setup requires installing Python and Visual Studio Code. Within VS Code, download the Azure Functions and Azure Account extensions from the marketplace. Create or sign into an Azure account through the Azure Functions interface, then generate a new function by selecting Python and your preferred environment. HOW-TO During an internship at Arinti, this guide was developed while combining Python with Azure to build a specialised application that converts sketches and electrical schema drawings into digital formats. This contribution comes from Cédric at Erasmushogeschool Brussel. #### Local installation and deployment caveat Once the project initialises, open the terminal and install your dependencies (e.g. pip install opencv-python numpy). This lets you use the Azure Function locally. However, there's a critical limitation: deployment to Azure may generate errors because packages install only locally, not on the Azure platform itself. Additional configuration is required for cloud deployment success. --- ### [Internship report: Can an AI play Rummikub?](https://arinti.ai/insights/internship-ai-play-rummikub) Category: article · Published 2021-04-01 > An internship project exploring whether AI can play Rummikub by combining computer vision with Monte Carlo Tree Search. From screen capture to board state The project used the official Rummikub app as a game platform, capturing screen sections with Python's Pillow library. OpenCV filtering extracted game elements and created contours to identify individual tiles. Two separate predictions occurred for each tile: digit recognition using a K-Nearest Neighbours algorithm (values 0-13, with 0 for jokers), and colour detection through RGB analysis using K-Means clustering. INTERNSHIP REPORT This internship project explores whether artificial intelligence can play the tile-based game Rummikub by combining computer vision with game algorithms. This contribution comes from Sander Backx at Thomas More Geel. #### Monte Carlo Tree Search for optimal moves Monte Carlo Tree Search simulated possible game states to determine optimal moves. The algorithm evaluated action sequences using a reward function: 100 points for player victory, move count plus 0.5 for ongoing games, and 0 for opponent victories. Once the board state was reconstructed from the vision pipeline, the algorithm could explore possible moves and select the most promising strategy. --- ### [AI and RPA for an automated sales order process](https://arinti.ai/insights/ai-rpa-automated-sales-order) Category: guide · Published 2021-04-01 > How AI and robotic process automation can automate repetitive tasks like sales order entry, demonstrated through an AI assistant called Raphael. An AI assistant for procurement Raphael is an AI-powered assistant that helps employees purchase everyday products for their organisation. It uses facial recognition for authentication and provides product recommendations based on previous orders. Once it gathers the necessary details through conversation, RPA technology automates the data entry into SAP systems. The combination works because AI handles the understanding — recognising faces, interpreting requests, making recommendations — while RPA handles the execution, entering data into enterprise systems exactly as a human would. AI & RPA How many times a day do you have to copy values from an Excel file into a form? Many jobs contain secondary, time-consuming tasks that have little to do with what you were actually hired for. AI and RPA can take those tasks off your plate. #### From order forms to document processing Beyond procurement, the AI + RPA combination addresses document structuring challenges across organisations. Language understanding models can reorganise unstructured documents, enabling RPA systems to process varied formats consistently. This was a joint demonstration between FLINQ and Arinti, showing how these two technologies complement each other in practice. --- ### [Microsoft Power Virtual Agents: a first look](https://arinti.ai/insights/microsoft-power-virtual-agents-first-look) Category: article · Published 2021-04-01 > A first look at Microsoft Power Virtual Agents: a no-code tool that lets subject matter experts build virtual agents without developers. No-code chatbot creation for business users The platform enables subject matter experts to create virtual agents without the need for data scientists or developers. Using a point-and-click interface, teams can build conversational agents that handle customer service questions and other types of external or internal inquiries. MICROSOFT Microsoft Power Virtual Agents empowers subject matter experts throughout organisations — including those in customer service, sales, marketing, finance, and HR — to develop virtual agents using a guided, no-code graphical interface. #### Freeing staff for higher-value work By enabling virtual agents to handle routine queries, organisations can redirect staff attention to more sophisticated responsibilities. Announced at MS Ignite, Power Virtual Agents represents Microsoft's push to democratise AI tooling across the enterprise. The Arinti team evaluated the platform and created a summary and demonstration of its capabilities. --- ### [Azure Machine Learning – The Responsible Road](https://arinti.ai/insights/azure-machine-learning-responsible-road) Category: article · Published 2021-04-01 > Exploring responsible machine learning through Microsoft's framework: understanding, protecting, and controlling AI systems. Interpretability and fairness Microsoft developed tools enabling interpretability of AI models. Interpret-Text provides explainable text-based models, FairLearn helps mitigate algorithmic bias, and DiCE offers counterfactual explanations — helping answer the question of what would need to change for a different outcome. RESPONSIBLE AI AI transforms sectors from medical diagnostics to autonomous vehicles. But with that power comes responsibility. Data can tell the truth, but it is not always the full truth — and algorithmic decisions increasingly affect people's lives, from loan approvals to parole determinations. #### Privacy-preserving analysis Security-focused tools address data privacy at the infrastructure level. Differential Privacy WhiteNoise enables privacy-preserving analysis, SEAL SDK supports homomorphic encryption, and Open Enclave provides a framework for processing encrypted data in secure environments. Tracking and auditing model iterations Azure Machine Learning Audit Trail tracks model iterations and dataset changes automatically. This is essential for compliance and accountability — knowing exactly which data and parameters produced a specific model version. Originally published on datafish.eu, this article summarises insights from a livestream discussion between four Microsoft AI MVPs. --- ### [Chatbot human handover with Microsoft Teams](https://arinti.ai/insights/chatbot-human-handover-microsoft-teams) Category: guide · Published 2021-04-01 > How Arinti enables seamless handover from chatbot to human agent through Microsoft Teams integration. Putting agents in the driver's seat when needed The integration maintains one-to-one communication scalability while allowing agents to humanise interactions and personalise customer experiences. The handover occurs in real time, ensuring smooth transitions without losing customers during difficult conversations. CHATBOTS Customers frequently ask questions that chatbots haven't been trained to handle. Without a fallback, this leads to frustration and repetitive unhelpful responses. Arinti built a feature that enables seamless handover of chatbot conversations to human agents through Microsoft Teams. #### AI efficiency meets human empathy This capability helps businesses maintain customer satisfaction by preventing scenarios where bots provide inadequate answers. The feature is designed to support both AI technology and human expertise working collaboratively — the bot handles volume and routine queries, while humans step in for nuance and complexity. --- ### [A pragmatic approach to identifying AI business cases](https://arinti.ai/insights/how-to-identify-ai-projects-business) Category: guide · Published 2021-04-01 > At Arinti, we use a structured data audit methodology to identify AI-powered business cases. Four steps, three deliverables, and a roadmap to get your AI journey started. Four steps to your AI roadmap Step 1: Kick-off session — a semi-structured interview to identify your 3-5 year strategic goals and brainstorm how AI can support them. Step 2: Business case identification — an analytical report mapping AI-powered business cases into a high-value, low-cost quadrant. Step 3: Data inspection — case-level assessment of required data sources, quality, and integration gaps. Step 4: Roadmap — a hands-on document outlining required steps, proposed technical architecture, and essential profiles for each business case. DATA AUDIT Our methodology for a data audit consists of four steps resulting in three key deliverables. The focus: small, short-term, low-risk projects close to your core business. This approach builds goodwill, identifies data management issues, and creates spillovers for future projects. Only target data connected to identified business cases. Nothing more, nothing less. --- ### [Four years Arinti – Four lessons from four years of AI projects](https://arinti.ai/insights/4-years-arinti-4-lessons-ai-success) Category: article · Published 2021-03-01 > After four years of AI projects, here are the four most important lessons we've learned: be open to change, empower everyone, account for hidden costs, and define your goals before you start. Be open to change — and communicate transparently Implementing AI always comes with change management. Transparency should be the top priority: explain what AI does, how it operates, how it's trained, and where it gets its information. At Unilever, we implemented AI using 10 years of data. Project lead Robin made it his duty to spread AI awareness within the company. AI STRATEGY An AI project always comes with a learning curve. Here are the four most important lessons we've learned to turn AI projects into successes. #### Account for hidden costs In 99% of AI projects, you'll pile on extra costs — mostly the time different internal stakeholders invest. IT provides system access, domain experts explain the business, ambassadors defend the project internally, and employees learn to work with AI. Time is your most precious resource. Empower everyone — inside and outside AI empowerment needs to be nurtured bottom-up. Get IT involved from the start — they hold the key to important data. Don't forget external stakeholders either. For Socialistische Mutualiteiten's chatbot for pregnant women, we organised workshops with moms-to-be to get their input. Define your goals before you start No goal = no viable AI project. Without an identified goal, projects rarely go into production. Partena Professional came to us with a clear goal: saving time and automating expert HR services. Less than three months after PoC, chatbot Louise became an essential tool for 900 payroll consultants. --- ### [Smart Monitoring: Predictive Maintenance and Anomaly Detection Explained](https://arinti.ai/insights/smart-monitoring-predictive-maintenance-anomaly-detection) Category: guide · Published 2021-01-15 > We built a smart monitoring dashboard for Fednot using Azure Databricks, Anomaly Detector API, and Power BI, turning unstructured log files into predictive maintenance insights. Converting unstructured data to structured data Log files are text-heavy and unstructured. The first step was extracting features — structured attributes close to the true factors for prediction. This requires deep understanding of both the data and the business problem. We created a transitional dashboard for feature exploration, letting Fednot select different feature combinations and trace back to original log entries — no coding required. PREDICTIVE MAINTENANCE Predictive maintenance reduces downtime, operational costs, and unscheduled maintenance while improving service quality. For Fednot (Royal Federation of Belgian Notaries), we built a smart monitoring dashboard with Azure and Power BI — not just as a proof of concept, but also to introduce their infrastructure team to different levels of Azure AI services. #### Azure AI Services across skill levels The data pipeline ran on Azure Databricks (code-heavy, for data experts). Anomaly detection used the Anomaly Detector API from Cognitive Services (low-code, a few lines of code). Visualisation used Power BI dashboards (no-code, for decision-makers). The Anomaly Detector takes any time-series dataset, automatically fits a model, and returns expected values, boundaries, and abnormalities — useful for both monitoring and predictive scaling. Results Even with limited effort, Fednot gained significant additional value from their monitoring data. The project is a first step towards a data-driven infrastructure environment: reducing unexpected downtime, enabling incident pattern discovery, and improving application quality and customer experience. — Fisher Kuan & Wouter Baetens --- ### [Smart Policing Hackathon: AI-driven crime investigation](https://arinti.ai/insights/smart-policing-hackathon) Category: article · Published 2020-12-15 > Arinti won the Smart Investigation challenge at the first Belgian Smart Policing Hackathon with Tafuta, our Cognitive Search and Graph Analytics solution for knowledge discovery in criminal data. The problem Huge amounts of information created daily are hardly used effectively. Entities like persons, vehicles, phone numbers, and known affiliations must be manually tagged in reports — extremely time-consuming. Analysts working on connections are scarce, and there's no proactive information sharing or concept-aware search. Data is siloed and its potential largely unexplored. COGNITIVE SEARCH From 9–11 December 2020, the Belgian Federal Police, VIAS and EY organised the first Smart Policing Hackathon. Arinti was selected as the winner of the Smart Investigation challenge with Tafuta, our Cognitive Search & Graph Analytics solution. The challenge: find innovative investigation techniques to improve the fight against serious and organised crime, specifically around knowledge discovery and connecting information across siloed data. #### Tafuta for law enforcement We prototyped a Police version of Tafuta that understands documents, interprets text, and turns raw information into searchable content. It recognises entities automatically and understands concept variations. We added speech-to-text for wire tap transcription, object detection for annotating images, and proactive alerts when new information appears about tracked entities. Graph analytics builds visual representations of how entities are connected — letting investigators literally see how the dots connect. The jury included the Commissioner General of the Federal Police and other high-level experts. --- ### [Data-driven companies: when being good enough is no longer sufficient](https://arinti.ai/insights/data-driven-companies) Category: article · Published 2020-11-01 > Data-driven companies outperform their competitors, yet 9 out of 10 firms point to cultural challenges as the biggest bottleneck. Three practical recommendations for becoming data-driven. Three practical recommendations Pick the low-hanging fruit first: focus on clearly identified high-impact problems with a critical business need. This builds value, credibility, and momentum. At Arinti, we use Data Audits and AI Project Canvases to uncover viable AI cases with interesting ROI. Let data flow freely through the organisation — it shouldn't be tucked away in a dark corner. And practice patience: becoming data-driven doesn't happen overnight. Don't abandon efforts when results aren't immediately impactful. DATA STRATEGY Data-driven companies outperform their competitors. Yet despite increasing investments in data and AI each year, companies still struggle to fully reap the benefits. 9 out of 10 firms point to cultural challenges — not technological ones — as the bottleneck. Organisational alignment, change management, people's skill sets, and resistance to change slow down the process. #### Mindset comes first At Arinti, we see first-hand that data science projects often require a cultural shift. The experimental nature of data science calls for agility and flexibility. It doesn't matter how big a company is or how long it's been around — mindset is the determining factor in whether a data science project will succeed. And not every experiment will succeed: there are always several ways of approaching a problem. We find the best solution and are not afraid to fail in the process. --- ### [Bachelor's thesis: Converting sign language into text using AI](https://arinti.ai/insights/sign-language-text-ai) Category: article · Published 2020-09-01 > We co-promoted a bachelor's thesis on using image recognition AI to convert Flemish Sign Language into written text, comparing Azure Custom Vision, Google AutoML Vision, and Amazon Rekognition. Technical approach Since video content recognition wasn't mature enough, the approach splits gesture video into frames and sends selected frames to an image classification API. Five frames per gesture are extracted for optimal performance vs accuracy. The thesis compared three services: Azure Custom Vision, Google AutoML Vision, and Amazon Rekognition Custom Labels. COMPUTER VISION Flemish Sign Language (VGT) is the mother tongue of approximately 6,000 Flemish people. Yet it's understood by only a small fraction of the population. This bachelor's thesis investigated whether AI can promote communication between sign language and spoken language by converting gestures into written text via camera. #### Training and results Using Azure Custom Vision, the model was trained with about 50 images per gesture as a starting point. Three factors determine quality: quantity, balance across labels, and variety (backgrounds, lighting, angles). A feedback mechanism lets users flag incorrect translations, enabling continuous improvement. The proof of concept confirmed that image recognition technology has advanced far enough to convert sign language into text. The reverse direction uses the VGT dictionary to display signs for spoken or typed input. Student: Yasmine De Winne — University College Ghent, co-promoted by Wouter Baetens --- ### [Internship report: K-Means clustering EWS data](https://arinti.ai/insights/internship-kmeans-clustering-ews) Category: article · Published 2020-07-01 > Intern Senne built a patient monitoring dashboard with K-Means clustering on Early Warning Score data, categorising patients by health trajectory to support clinical decision-making. From exploration to clustering After data exploration in Python — visualising distributions, filtering healthy vs unhealthy data, and identifying boundary zones — the project moved to K-Means clustering. The algorithm took patients with 100+ measurements, extracted their 87 most recent EWS values, and assigned cluster labels based on health trajectory trends. INTERNSHIP During an internship at Arinti, Senne (Erasmushogeschool Brussel) worked with an anonymised dataset of over 170,000 Early Warning Score (EWS) measurements. The EWS score aggregates vital parameters like heart rate, blood pressure, and body temperature to indicate a patient's overall condition. #### Dashboard and results An Angular web dashboard displayed monitoring data per patient, with colour-coded values for danger zones. The clustering results showed each patient's category with an explanation of what it means for their health outlook. This can help doctors and nursing staff assess patient condition and predict future trajectory. — Senne, Erasmushogeschool Brussel --- ### [Dealing with Optimisation Problems: FitJSP](https://arinti.ai/insights/optimisation-problems-fitjsp) Category: guide · Published 2020-05-01 > How we used genetic algorithms to solve a job-shop scheduling problem and built a drag-and-drop online tool for interactive schedule management. Genetic algorithm approach The genetic algorithm starts with a random initial population satisfying machine constraints. A fitness function scores each individual based on makespan, daily utilisation, and conflict avoidance. Crossover creates children that take the best rows from each parent (minimum collisions), while mutation stochastically shuffles timeslots and machines to explore alternative paths. The process iterates until the desired performance score is reached or no further improvement is found. OPTIMIZATION Job-shop scheduling is a well-known optimisation problem: finding the optimal strategy to distribute multiple tasks over several available machines. Our specific problem involved in products requiring both a preparation and drying process, each with defined time requirements and machine capacity constraints, scheduled on a weekly basis. #### Interactive drag-and-drop tool On top of the algorithm, we built FitJSP — a drag-and-drop online tool where users can interact with the recommended schedule in real time. Users can update, move, create, delete, or resize events. A notification panel informs users in real time whether changes create conflicts. This makes the output accessible to non-technical users. — Ali Doku --- ### [Build your own image dataset with Bing Image Search API](https://arinti.ai/insights/data-gathering-image-recognition) Category: guide · Published 2020-03-01 > A step-by-step Python guide for building your own image dataset using the Bing Image Search API, from search and download to storing results in Azure Machine Learning Studio. The workflow In Python: install the Bing Image Search SDK, configure your subscription key and endpoint, search with your keyword and desired count, convert results to a Pandas DataFrame, generate unique filenames using MIME type detection and UUIDs, then download images with error handling for broken links. The whole script can be parametrised to accept an array of search terms for batch downloading. TUTORIAL Need training data for image classification but can't find it on Kaggle? The Bing Image Search API (part of Azure Cognitive Services) lets you search for images programmatically, filtering by size, file size, licence, and colour. #### Moving data to Azure ML Studio If you're using Azure Machine Learning Studio, upload the image folder to a datastore so other data scientists can access it too. A few lines of code with the azureml.core library handle the upload. The full Jupyter Notebook is available on GitHub. --- ### [Alterations – A Hidden Gem In QnA Maker](https://arinti.ai/insights/alterations-hidden-gem-qna-maker) Category: guide · Published 2020-01-15 > QnA Maker's alterations feature lets you define synonym lists via the API, reducing training effort and improving accuracy, especially for non-English languages with fewer pre-trained synonyms. Why alterations matter for non-English bots QnA Maker is pre-trained with English synonyms. For other languages, synonym coverage is limited. Adding well-chosen alterations — for example, mapping 'GDPR' to 'AVG' in Dutch, or 'auto' to 'wagen' — significantly reduces the number of alternative questions you need to add to your knowledge base. How to add them via the API Use the REST API PUT endpoint at {Endpoint}/qnamaker/v4.0/alterations with your Ocp-Apim-Subscription-Key. The body contains a wordAlterations array of synonym groups. Note: this replaces all existing alterations and applies across all knowledge bases in your QnA Maker service. Note - QnA Maker retired on 31 October 2025 QnA Maker was retired on 31 October 2025. Its successor is custom question answering in Azure AI Language. Synonyms (the alterations equivalent) work much the same way — the principles in this article still hold. CONVERSATIONAL AI QnA Maker has many features: Active Learning, Multi Turn Questions, and one-click bot publishing. But there's a hidden feature available only through the API: Alterations. These are synonym lists that make your QnA Maker smarter with less training. --- ### [3 Different Flavors For Building Chatbots With Microsoft](https://arinti.ai/insights/3-flavors-building-chatbots-microsoft) Category: guide · Published 2019-12-01 > Microsoft offers three tools for building chatbots: Power Virtual Agent (no-code), Bot Framework Composer (low-code), and Bot Framework SDK (full flexibility). Here's how they compare and who each one is for. Power Virtual Agent — no code Built for business users. An online application with a clean GUI, pre-built topics, and no code required. It supports topic and entity recognition, Power Automate integration, human handover, and easy multi-channel deployment. The trade-off: no LUIS integration (entities must be asked for rather than recognised), no QnA Maker support, no language generation, no versioning, and no debugging. CONVERSATIONAL AI Microsoft now offers a chatbot building tool for everyone. At Ignite 2019, two new tools were announced alongside the existing SDK. Here's how they compare. #### Bot Framework Composer — low code The missing link for Microsoft Bot Framework. A visual dialog builder that integrates LUIS and QnA Maker training directly in the tool. It's fully open-source, so you can white-label it or build custom connectors. Having coding skills helps for extra features, but it's not required for core functionality. The tool is linked to version control, though multi-person collaboration can be tricky. Bot Framework SDK — full flexibility The highly adaptable SDK (v4.6+), available for C#, Java, Node.js, and Python. The sky is the limit: write your own integrations, use Adaptive Dialogs for context-switching mid-conversation, debug line by line in your IDE, and collaborate with other developers via version control. The learning curve is steep, but it gives full control. Start with Power Virtual Agent for simple bots. Move to Composer or SDK when you need more intelligence and flexibility. --- ### [Network flow: a powerful tool for modelling problems](https://arinti.ai/insights/network-flows-applicability) Category: article · Published 2019-11-01 > An introduction to network flow algorithms: from finding edge-disjoint paths to solving employee-project matching problems with bipartite graphs and max-flow-min-cost optimisation. From unique paths to bipartite matching Given a graph, finding unique (edge-disjoint) paths from node i to j is solved by setting each edge capacity to 1 and running maximum flow. For node-disjoint paths, duplicate every node into k and k' with capacity-1 edges between them. For employee-project matching: connect source to employees, employees to matching projects (based on skill requirements), and projects to sink. Set all edges to capacity 1. Maximum flow gives the optimal matching. ALGORITHMS Network flow is part of the graph theory toolbox, used to model transportation networks, scheduling, and resource matching. With today's libraries, a simple method call solves the optimisation — the hard part is modelling the problem correctly. #### Extending with real-world constraints Real problems have additional constraints: employee seniority ranks, project difficulty levels, budget limits. We modelled this with edge capacities equal to wages, source-to-worker edges based on simultaneous project capacity, and project-to-sink edges capped at project budget. Running max-flow-min-cost with networkx and filtering unsaturated edges gives budget-respecting assignments. — Armando ## Open positions ### [AI Engineer](https://arinti.ai/careers/ai-engineer) > Good AI doesn't stop at the model. As an AI Engineer, you design, build, and integrate AI-powered solutions grounded in what really works. You collaborate with colleagues across disciplines to address real business needs — not technology for technology's sake. Your mission as an AI Engineer Good AI doesn’t stop at the model. As an AI Engineer, you help design, build, and integrate AI-powered solutions, informed by the latest developments and grounded in what really works. Together with colleagues from different disciplines, you design and implement solutions that address the actual business needs of our clients, not technology for technology’s sake. What you’ll do - Design and build AI solutions end-to-end, from concept to production. - Design and implement agentic systems and workflows. - Collaborate with colleagues and client teams across roles and capabilities. - Help operationalise and scale AI solutions. - Contribute to our internal AI roadmap, research, knowledge sharing, and evolving our best practices and way of working. What you bring - 3+ years of experience in machine learning and software development. - Strong Python skills and solid software engineering principles. - A genuine interest in GenAI and agentic solutions. - Knowledge of MLOps principles and how to scale AI solutions. - Familiarity with cloud platforms and services. - A degree in Computer Science or a related field. - A proactive mindset, you work well both autonomously and as part of a team. - Fluent in Dutch and English. What we offer Experience a human-sized company with a startup culture and the professionalism of a larger organisation when it counts. Our company culture is transparent and pragmatic. Not only towards our clients, but also towards our people. We believe honesty is the best policy and we try to maintain an open mind in everything we do. - A competitive salary package, company car, and additional benefits tailored to your situation. - An environment with challenging projects that stimulate and support your personal growth. - A great team of like-minded people where drive, honesty, and transparency come first. - Room to shape your role as you grow with the team. JOB DESCRIPTION #### Build AI solutions with real impact, for clients that matter We’re an AI-native organisation with a track record that goes back well before the recent surge in GenAI. That experience shows in our broad capabilities, from Data Engineering to agentic AI solutions. You’ll work alongside a team that’s driven to find solutions, stays current with the latest technology, and doesn’t get lost in hype. CULTURE #### How we work at Arinti Experience a human-sized company — small enough that your work matters, structured enough that it ships. Our culture is built on transparency and pragmatism, towards our clients and towards each other. No politics, no layers of abstraction, just honest conversations about what works and what doesn’t. We’re a group of like-minded people devoted to technology, working at the frontier of data and AI. We look beyond the hype to deliver pragmatic solutions built on what really works. Simple, honest, very Arinti. --- ### [Data Engineer](https://arinti.ai/careers/data-engineer) > High-quality data first, smart processes next. As a Data Engineer, you help our clients design and implement robust, scalable data solutions and pipelines. You work on large-scale, challenging projects for clients that matter. You make sure their AI initiatives are built on solid data foundations. Your mission as a Data Engineer High-quality data first, smart processes next. As a Data Engineer, you help our clients design and implement robust, scalable data solutions and pipelines. You work on large-scale, challenging projects for clients that matter. You make sure their AI initiatives are built on solid data foundations. What you’ll do - Understand and translate our clients’ needs and business cases into robust data solutions. - Design and implement complex data solutions end-to-end, from data ingestion to production-ready platforms. - Collaborate in multidisciplinary teams to deliver scalable and reliable solutions. - Design, build, and optimise data lakehouses and warehouses. - Build scalable, cloud-based data platforms using major cloud providers. What you bring - 5+ years of experience in Data Engineering or comparable roles. - A degree in Computer Science or a related field. - Proven expertise in designing and building performant data pipelines and platforms. - Experience with data platforms such as Databricks, Snowflake, Fabric. - The ability to work with non-technical stakeholders, translating their goals into concrete technical solutions. - A good understanding of data governance principles and data quality frameworks. - Experience within major cloud ecosystems. - Fluent in English. Dutch or French is a strong plus. What we offer Experience a human-sized company with a startup culture and the professionalism of a larger organisation when it counts. Our company culture is transparent and pragmatic. Not only towards our clients, but also towards our people. We believe honesty is the best policy and we try to maintain an open mind in everything we do. - A competitive salary package, company car, and additional benefits tailored to your situation. - An environment with challenging projects that stimulate and support your personal growth. - A great team of like-minded people where drive, honesty, and transparency come first. - Room to shape your role as you grow with the team. Job description #### Build AI solutions with real impact, for clients that matter We’re an AI-native organisation with a track record that goes back well before the recent surge in GenAI. That experience shows in our broad capabilities, from Data Engineering to agentic AI solutions. You’ll work alongside a team that’s driven to find solutions, stays current with the latest technology, and doesn’t get lost in hype. CULTURE #### How we work at Arinti Experience a human-sized company — small enough that your work matters, structured enough that it ships. Our culture is built on transparency and pragmatism, towards our clients and towards each other. No politics, no layers of abstraction, just honest conversations about what works and what doesn’t. We’re a group of like-minded people devoted to technology, working at the frontier of data and AI. We look beyond the hype to deliver pragmatic solutions built on what really works. Simple, honest, very Arinti. --- ### [Data Architect / Lead Data Engineer](https://arinti.ai/careers/data-architect-lead-data-engineer) > We're an AI-native organisation with a track record that predates the recent GenAI surge — from Data Engineering to agentic AI. You'll work with a team driven to find solutions, current with the latest technology, and grounded beyond the hype. Your mission as a Data Architect / Lead Data Engineer Our enterprise clients trust us to design and deliver data platforms that scale, perform, and stand the test of time. As an experienced Data Architect and Engineer, you make that possible. You translate our clients’ strategy into concrete data solutions and architectural principles, then drive the implementation and coach less experienced colleagues in upholding those principles. You’re equally comfortable whiteboarding a lakehouse architecture with a CTO as you are rolling up your sleeves to troubleshoot a pipeline in production. What you’ll do - Understand the strategic drivers and business cases of our clients and translate them into concrete data solutions with the right architectural principles, requirements, and constraints. - Design and build complex data solutions and architectures end-to-end, from data ingestion to production-ready platforms. - Guide and support less experienced colleagues during implementation, and steer the technical execution where it matters. - Communicate technically complex concepts clearly, acting as a trusted technical advisor for client stakeholders and colleagues alike. - Contribute to our collective knowledge, share lessons learned and best practices, and help shape how we tackle data architecture at Arinti. What you bring - Significant experience (typically 7+ years) in data engineering and comparable roles. - Proven expertise in architecture and technical design of large-scale data platforms. - Hands-on experience with Databricks and ideally the Azure data ecosystem. - The ability to understand client strategy and translate business cases into concrete technical solutions. - Experience working in complex, large-scale environments and client-facing situations. - A solid understanding of data governance principles and data quality frameworks. - Experience leading and mentoring data engineers. - Experience within major cloud ecosystems (Azure, AWS, or GCP). - Fluent in English. Dutch or French is a strong plus. What we offer Experience a human-sized company with a startup culture and the professionalism of a larger organisation when it counts. Our company culture is transparent and pragmatic. Not only towards our clients, but also towards our people. We believe honesty is the best policy and we try to maintain an open mind in everything we do. - A competitive salary package, company car, and additional benefits tailored to your situation. - An environment with challenging projects that stimulate and support your personal growth. - A great team of like-minded people where drive, honesty, and transparency come first. - Room to shape your role as you grow with the team. Job description #### Build AI solutions with real impact, for clients that matter We’re an AI-native organisation with a track record that goes back well before the recent surge in GenAI. That experience shows in our broad capabilities, from Data Engineering to agentic AI solutions. You’ll work alongside a team that’s driven to find solutions, stays current with the latest technology, and doesn’t get lost in hype. CULTURE #### How we work at Arinti Experience a human-sized company — small enough that your work matters, structured enough that it ships. Our culture is built on transparency and pragmatism, towards our clients and towards each other. No politics, no layers of abstraction, just honest conversations about what works and what doesn’t. We’re a group of like-minded people devoted to technology, working at the frontier of data and AI. We look beyond the hype to deliver pragmatic solutions built on what really works. Simple, honest, very Arinti.