The situation
Listing prices vary wildly between portals, the same flat appears three times under different names, and neighbourhood names are spelled ten ways. A model trained on that raw mess learns the mess.
Data and ML · Real estate
A price model built from 71,000 real listings
Advisors needed a fair price estimate before visiting a property. We built the whole chain: scrapers across several property portals, a cleaning and deduplication pipeline, and a model served as an API that answers with a range and an honest reliability rating.
Illustration
Data and ML · Real estate
- Collect
- Clean and deduplicate
- Place on the map
- Train
- Serve the estimate
What changed
The pipeline collects, cleans and deduplicates listings across sources and maps them to a curated list of neighbourhoods. The model returns a price range, a probable sale price after negotiation, and a reliability rating that drops where data is thin.
How it works
5 steps, end to end
- 01
Collect
Scrapers gather listings from 7 sources across several property portals.
- 02
Clean and deduplicate
Exact and cross-source duplicates are removed; prices and surfaces are normalised.
- 03
Place on the map
Each listing is matched to a curated gazetteer of 410 neighbourhoods in 6 cities.
- 04
Train
A gradient-boosted model learns from the 22,371 listings kept.
- 05
Serve the estimate
A stateless API returns a range, a likely sale price and a reliability rating.
Where the AI stops
The model writes. The code decides.
The model
- Learns price patterns from cleaned listings
- Estimates a range and a likely sale price
The code
- Deduplicates across portals before training
- Maps every listing to a known neighbourhood
- Lowers the reliability rating where data is thin
By the numbers
Counts read in the code we shipped. No invented results.
- ~71k
- listings collected
- 22,371
- kept after cleaning
- 410
- neighbourhoods in the gazetteer
- 7
- sources scraped
Built with
- Web scraping
- Data cleaning and dedup
- Gradient boosting (CatBoost)
- Stateless price API
Built for
The client behind it
Real estateSaafkaA network of real-estate advisors: buy, sell, rent and appraise anywhere in Morocco.Read the client storyMore use cases
All use cases
- AI agentsA legal assistant that cites the law, or says it doesn't knowA legal assistant that only answers from the law it retrieved, and checks every article it cites.SAAFKA
- Voice AIA sales trainer that picks up the phoneTrainees phone a realistic AI property owner, then get a scored debrief with one action to work on.ConciergElite
- AI agentsA coach that reads the playbook so advisors don't have toA coaching agent that retrieves only the right sections of a large knowledge base, and keeps internal names out of its answers.SAAFKA
Contact
let's talk about your project
- Where we areMorocco, working worldwide
- Writehello@kaishift.com
- LanguagesFrench, Arabic, English
Write to usWe answer the same working day.