Darkom & Avito
End-to-end Moroccan real-estate pipelines feeding a price-prediction ML service and BI dashboards.
Voir le codeProblème
Moroccan real-estate listings are messy, duplicated and inconsistently priced — making it hard to trust any price estimate or market trend drawn from them.
Approche
Extracted listings via Selenium into a PostgreSQL staging layer, modeled a Snowflake star schema (fact_annonces + 12 dimensions), then trained scikit-learn and XGBoost regression/classification models for price prediction and price-category tagging — served through FastAPI with SHAP explainability, DVC-tracked data versions, and full observability via MLflow.
Défi
Listing data needed careful outlier handling (IQR-based flags), Moroccan city-tier classification, and a production-grade test suite before the model could be trusted — reaching 254 passing tests with zero failures.
Résultat
A deployable, explainable price-prediction API backed by a proper data warehouse and Power BI dashboards for volume, price trend and segmentation.