Three out of four people selling their house define it as a “stressful” process, whether it’s about defining the right price, listing your property or actually moving out – it’s a big task up front and even bigger halfway through. Finding the right partner in doing so can relieve much of that stress and make selling or renting property a much more enjoyable experience.

That is, if you can find the right partner. Smooved set out to make that search easier to a large degree by building HouseMatch, where potential sellers or landlords can scan the area around their property and get recommendations for highly professional brokers nearby.

Making those suggestions is what we do with our team: collecting heaps of data of these brokers and selecting the best matches is what we contribute to this – real time on the platform as we speak. Successfully, too, generating visits to almost a fourth of the Belgian broker market on launch day, which lead to over 100 new brokers looking to boot up a partnership with Smooved as well. Talk about utilising #FOMO to drive business towards our partners while helping resolve stress, it’s a win in our books!

It made immediate sense what “success” meant in Smooved’s books: generating contact requests between property owners and broker groups by making the ideal recommendations. In hindsight, we should have focused more on that statement than we did at the start: Smooved already had a good idea of what made a recommendation “ideal”.

However, in our eagerness to get started, we rolled up our sleeves and started collecting valuable data to drive several potentially successful machine learning algorithms that would provide those links between owner and broker. Collecting data is what we do best, and our research and selection of both internal and external data sources proved valuable to both our algorithms and our dataset. We ingested hundreds of thousands of records and generated a pretty good view of what brokers were up to in all regions across Belgium. On that dataset, we trained multiple models to figure out what parameters drove the best recommendations, but kept short of making this explainable to both Smooved and end-users: quite the mistake.

After a couple of sessions trying to get everyone on board of seemingly productive models, we ground to halt. We were going the wrong way – our model wasn’t explaining any of it’s decisions, and the Smooved sales team needed that information to keep brokers on board of the tool. Both Smooved and our team were unhappy, and for a moment things seemed hopeless in terms of delivery in time.

However, giving up is wasted on us; with renewed enthusiasm we got back to the basics and sat down with Smooved once more, but this time with an entirely open mind. We worked hard and fast on what our model should take into account and what was necessary to explain to others, defining the musts and killing our won’ts. We upgraded our dataset to match the required input and created a custom model: better, faster and more importantly exactly as Smooved expected.

Sprint after sprint we refined the system, collecting output and validating results in in-person workshops, combining technical and business insights from the team, weeding out edge cases and honing the parameters until they fit perfectly – right on time, too. A few weeks before the launch of HouseMatch we got a production-ready API, integrated in the tool and only susceptible to a few small tweaks still to be made.

We’ve gotten to the end; a recommendation engine with more than sufficient data and insight to suggest brokers to real estate owners that made Smooved feel comfortable, excited even, to launch. That’s where we are today – a comfortable collaboration between our teams with monitoring and upkeep of the engine, to ensure a constant correct working of the entire HouseMatch application.

So what exactly did we build? Without getting too much into detail (get in touch if you want to know more): a large data warehouse and a custom-made algorithm.

The data warehouse

Setting up a data warehouse with over potentially millions of records require preparation, a good data model and a cloud-first approach. As a Microsoft Gold Partner, it’s easy to imagine us taking to Azure to set up our warehouse using our very own bootstrap framework to gain traction (and reduce implementation costs). This cloud solution allows for fast scaling and accurate hosting costs – you only pay for what you use, when you use it.

We set up the traditional four-layer structure (from ‘raw’ to ‘gold’) and ingested both Smooved data using internal API’s, as well as collecting broker info and reviews from Google and Facebook places with external API tools. This data is collected and harmonised throughout the layers, and enriched with additional content in the gold layer. For example: we automatically scan all incoming reviews with a language model to detect best and worst scoring brokers services to get an overal view of how well brokers are performing. This way, we can easily label over 5.000 brokers and suggest various actions on their account to score better in recommendations (and actual business), all without lifting a finger.

(This works on any text-like input too: scanning incoming mails to create tasks or send automated responses, to having a conversation with support users without intervention from support officers.)

The algorithm

Step two is using that data (with internal collections) to predict the best brokers nearby, based on input from HouseMatch customers in the interface. Explaining the entire algorithm would compromise IP, we can only tell you lots of standard deviations were involved, and a mathematical (very explainable) model was created to fit both end-users and Smooved’s needs.

The interface links to said model using a secure endpoint, and within seconds a set of recommendations are presented. We’ve versioned that endpoint, too, to be able to compare performance and recommendations, selecting the right version for launch and continuing our work in the background.

Additionally, to identify feedback and sentiment on specific services from brokers in reviews, we use a Zero-Shot Learning language model (a BERT model) to limit the need for training data (which meant gaining months in terms of using the model and endless hours saved on creating samples). With accuracy on sentiment scoring just over 95%, it was the obvious winner in our implementation, so we trained it with additional specific keywords and got to where we are today: an automated pipeline telling us which brokers are performing which great services (and wherein they are worse).

Our journey with Smooved already proved valuable: we’re connecting people to the right brokers, relieving stress in selling or renting property and providing insights to the real estate scene in Belgium. It also showed us there’s so much more to do still, such learnings hidden in the data we currently ingest and analyse, so much more value to create and deliver to end-users. Nathan says it best:

“We put our eyes on a target and keep grinding – day after day.”

We’re continuing our work with Smooved and HouseMatch, learning about the best way to present data, recommendations and business value along the way, so join us in looking forward to our extended partnership and the interesting projects that will result from it!