SafwenCherif/tunisian-real-estate-mlops

# Tunisian Real Estate MLOps — Apartment Price Prediction > A fully automated, production-grade MLOps pipeline that wakes up every morning at 03:00 UTC, checks whether the Tunisian real estate market has new listings, scrapes only what is new, enriches it with geographic data, retrains a price-prediction model if the data changed, promotes the best model to production, and backs everything up to Google Drive — with no human intervention required. **Data source:** mubawab.tn **Target:** Apartment sale prices in TND (Tunisian Dinar) **Champion model:** Ridge Regression — R² ≈ 0.77 · average error ±22% (Can be changed if the data changes) **Stack:** Python · Airflow · DVC · MLflow · Docker · Scikit-learn · BeautifulSoup · Geopy · Nominatim --- ## Table of contents 1. Project overview 2. Repository structure 3. Architecture — the four layers 4. Infrastructure layer — Docker 5. Scheduling layer — Apache Airflow 6. ML reproducibility layer — DVC 7. Experiment tracking layer — MLflow 8. Pipeline scripts — role of every file 9. Notebooks — EDA and modeling in detail 10. Airflow task logs — how output flows 11. Data artifact lineage 12. Complete execution timeline 13. Key design decisions and why 14. How to run the project 15. Project results summary --- ## 1. Project overview This project predicts the sale price of Tunisian apartments from data scraped from mubawab.tn. It is not just a machine learning notebook — it is a complete MLOps system built on four distinct, non-overlapping layers that automate the entire lifecycle from raw web data to a promoted production model. The pipeline runs on a fixed daily schedule. On each run it: 1. Checks whether Mubawab has new apartment listings since the last run 2. Scrapes only the new listings using a content-based deduplication fingerprint 3. Geocodes only the new rows using the Nominatim API and computes 14 geographic features per row 4. Runs the EDA and modeling notebooks headlessly via `nbconvert` 5. Compares the newly …