Production ML and reliability

This path follows a model and its data into operational conditions: temporal evaluation, typed serving, model promotion, and segmented drift detection. All core exercises run offline; MLflow and serving integrations are optional.

Time	Extras	Output
45–90 min	ml	PIT dataset, metrics, model artifact
30–60 min	serving, mlflow (optional)	service checks, promotion manifest
30–60 min	base + ml (optional)	drift report and segment alerts

The model lifecycle notebooks include optional mlflow and FastAPI/Uvicorn client cells. They log or score a bounded artifact when available and retain a local manifest fallback when no tracking server or HTTP process is running.

Tutorials

• Build a Simple Fraud Scoring Model

• Train, evaluate, and track a fraud model

• Promote and serve a model locally

• Resume a scale run from a checkpoint

• Promote, reject, and roll back model versions

• Measure drift by operational segment

Next path: Graph analytics.