Smooved - Recommending brokers to house owners
Belgium has over 5,000 active real estate brokers. For property owners looking to sell or rent, there was no data-driven way to compare broker performance and find the right match for their area. Together with Smooved, we built a recommendation engine for HouseMatch.be that analyses broker data daily and surfaces ranked suggestions based on objective, explainable criteria.
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.
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.
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.
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.
What was built
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.
Impact at launch
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.”
