Real Estate ML Microservice

Overview

A production-grade Python microservice that exposes property valuation predictions via a REST API, trained on historical transaction records, macroeconomic time-series data, and geospatial features derived from OpenStreetMap and government land-use datasets.

The modelling pipeline uses a gradient-boosted ensemble (XGBoost + LightGBM) with feature engineering covering property attributes (size, age, floor, facing direction), neighbourhood metrics (average transacted prices, distance to MRT/schools/malls), and macro factors (interest rates, CPI, HDB resale index).

The training pipeline is orchestrated with Prefect, pulling fresh data from the data warehouse nightly, retraining models, and promoting them to production if they beat a held-out validation benchmark. Model artefacts are versioned with MLflow.

The FastAPI service wraps the loaded model for real-time inference, returning a valuation estimate with a confidence interval and top feature contributions (via SHAP values) so users understand what drives the prediction. A companion batch scoring endpoint processes bulk property lists for portfolio analysis. Prediction latency averages under 50 ms, and the model achieves a median absolute percentage error of under 6% on out-of-sample test transactions.