lolia592/olist-late-delivery ? reverse-engineered prompt

Build me a small production style app that predicts whether an Olist e commerce order will be delivered late or on time, using the order, payment, and customer data from the notebooks.

I want the project to include a FastAPI service with a health check, model info, single prediction, batch prediction, and metrics endpoints, plus a clean inference pipeline that reads everything from config instead of hardcoding paths. Please make the input validation strict, so bad requests are rejected with clear 422 errors, and deeper data quality issues are caught before prediction with helpful messages. The model should load from MLflow registry style artifacts and behave the same way as the notebook output, including the known quirk that the saved scaler is not used in inference.

Also set it up so I can run the whole thing locally with Docker Compose, run tests with pytest, and have a simple one command health check script. If you need current docs for FastAPI, MLflow, or Docker details, look them up online.