Customer Churn MLOps

The project treats customer-churn prediction as an operational lifecycle rather than only a training script. It combines data validation, drift reporting, model training and registration, controlled promotion, and API serving.

Incoming CSV files are validated against a contract and routed to accepted or rejected datasets. Feature and target drift reports determine whether the lifecycle stops after reporting or continues to dataset curation, candidate training, model selection, and MLflow/DagsHub registration.

DVC versions pipeline inputs and generated artifacts, while a fixed test dataset and API-compatibility checks are part of the gates before a candidate is eligible for manual promotion to the champion alias.

An optional agent can propose bounded experiments and write an analysis report. Deterministic code validates data and plans, trains candidates, selects the model, registers it, and evaluates promotion gates; the LLM does not make operational decisions.

The same lifecycle can run locally or through GitHub Actions. After promotion, a FastAPI service can serve the model registered under the champion alias through health, model-information, and prediction endpoints.