A Production-Ready MLOps Framework for Customer Churn Prediction with Explainability, Drift Detection, and Fairness Auditing

This work presents a production-ready MLOps framework for customer churn prediction that integrates model training, explainability, distribution drift detection, fairness auditing, experiment tracking, model serving, and monitoring into a reproducible machine learning pipeline. The framework benchmarks Random Forest, Gradient Boosting, XGBoost, LightGBM, and CatBoost under multiple imbalance-handling strategies. CatBoost with threshold tuning achieved the best overall performance with an AUC-ROC of 0.953 and PR-AUC of 0.923. SHAP-based explainability identifies customer service calls, international plan subscription, and daytime usage intensity as the strongest churn drivers. The framework further evaluates drift detection using Evidently AI and PSI, performs subgroup fairness auditing using Fairlearn, and demonstrates a complete deployment architecture using MLflow, DVC, FastAPI, Docker, Grafana, and GitHub Actions. The repository and accompanying artifacts are released to support reproducible research, educational use, and future extensions in customer analytics and MLOps.