Explainable Churn Prediction Workbench for Telecom Retention Teams

Telecom operators lose a meaningful share of subscribers every quarter, and retention teams cannot act on a churn score alone — they need to know why a specific customer is flagged so they can choose the right offer or intervention. This project builds a full-stack application that pairs accurate churn prediction with human-readable explanations. The intern explores and prepares subscriber data with pandas and NumPy in Jupyter, trains an XGBoost classifier with scikit-learn tooling for validation, and applies SHAP to attribute each prediction to concrete drivers such as contract type, tenure, or support call history. A FastAPI backend serves the model behind clean prediction and explanation endpoints, while a Streamlit interface lets analysts upload customer records, review churn probabilities, and inspect force plots and feature importance charts interactively. The application is containerized with Docker so the full stack can be deployed as a single unit. By finishing the project the intern demonstrates the complete applied machine learning workflow — data preparation, model training, interpretability, API design, and deployment — and shows they can turn a black-box model into a decision-support tool a business team would actually use.