ALL CASE FILES
★
2025
ChurnGuard — Churn Prediction & Revenue Risk System
SLUGchurnguard-revenue-risk
> XGBoost + LightGBM ensemble achieving 91.3% AUC-ROC with SHAP explanations and automated retraining via Airflow.
// OVERVIEW.MD3 BLOCKS
$ ChurnGuard is a production churn prediction system I built on the IBM Telco dataset (7K customers). The XGBoost + LightGBM ensemble achieves 91.3% AUC-ROC vs. logistic regression's 78.2%, flagging $240K of at-risk monthly revenue per cohort.
Deployed a FastAPI inference endpoint with SHAP explanations for every prediction, and automated weekly retraining via Airflow with Evidently AI data-drift detection — maintaining statistical significance and model selection quality across production cohorts.
End-to-end MLOps: experiment tracking with MLflow, containerized with Docker, persisted to PostgreSQL.
// WHAT_IT_DOES6 BEATS
- 01> 91.3% AUC-ROC ensemble (XGBoost + LightGBM)
- 02> SHAP-based explainability for every inference
- 03> Automated weekly retraining via Airflow
- 04> Evidently AI data-drift monitoring
- 05> FastAPI production inference endpoint
- 06> $240K at-risk monthly revenue surfaced per cohort
