Challenge
PropConnect was aggregating listings from 12 provincial MLS boards but facing low engagement. Buyers were overwhelmed by 850K+ listings with basic filtering, and sellers couldn't understand why their properties weren't converting.
Solution
Mapletech Labs built an ML-powered recommendation engine that predicts property appeal for each buyer, personalizes search results using behavioral data, and surfaces high-conversion listings first.
Results
Buyer conversion rate jumped 35%, time-to-conversion decreased 28%, and seller inquiries increased 42%. The platform now handles 2.4M monthly visitors with 120ms search latency using Elasticsearch and TensorFlow.
The Challenge
Real Estate Search That Actually Works
PropConnect Canada aggregated real estate listings from 12 provincial MLS boards across Canada — a massive dataset of 850K+ active listings. But having the largest listing database meant nothing if buyers couldn't find what they wanted. Users were overwhelmed by choice paralysis, search results were sorted by price or date (generic and ineffective), and buyers were spending 3-4 hours on the platform to find 5 relevant properties.
Sellers were frustrated too. A $650K home in Toronto would sit for months while an inferior property sold in weeks. PropConnect had no way to surface high-quality properties to the right buyers, and sellers blamed the platform instead of blaming the actual market dynamics. Real estate agents were switching to Zillow and Realtor.com because they had better recommendations.
PropConnect needed to inject intelligence into search — using data on user behavior (which properties they clicked, which they saved, search history, time spent per listing) to predict what buyers actually wanted, then personalize results for each visitor. This meant building a real-time ML recommendation system that could score 850K+ properties in <200ms.
Our Solution
ML Recommendations with Real-Time Search
We built a two-tier recommendation system. First, Elasticsearch indexes all 850K listings with metadata (price, location, sqft, property type, age, school district, walkability score). When a user searches, the Node.js API queries Elasticsearch for candidate properties (120ms with proper indexing). Second, a TensorFlow model trained on 2+ years of PropConnect user behavior (click data, saves, time spent, conversions) scores each candidate property in real-time for that user.
The ML model uses collaborative filtering (if you liked homes similar to X, you'll like X) combined with content-based filtering (properties similar to your saved favorites) plus contextual signals (distance to work, school ratings, crime data). The model processes each user's behavioral profile (searchHistory, saveHistory, timeSpentPerProperty, dayOfWeek, timeOfDay) to personalize results. Results are re-ranked by predicted likelihood of conversion, not price or date posted.
The backend is Node.js + GraphQL exposing properties, user preferences, and recommendations. React frontend displays personalized results with explanations ("We think you'll like this because it's similar to X and Y"). We built automated feedback loops — every search, click, save, and listing view trains the model incrementally. Real estate agents get a dashboard showing "seller insights" (e.g., "Your property matches 2,400 buyers in your price range, average time-to-view: 3 days") — this transparency increased seller satisfaction significantly.
Key Results
Turning Data into Real Estate Gold
“The ML recommendations changed everything. Our sellers are happier because their homes get in front of the right buyers. Our buyers find homes faster. Conversion rate jumped from 22% to 35% — that's a multi-million-dollar impact. Mapletech Labs didn't just give us a technical solution, they gave us a competitive advantage that's defensible and only gets stronger with more data.”
CEO
PropConnect Canada