Customer Churn Prediction

The idea behind it.

This project is a full-stack AI product for predicting customer churn. A Flask API serves a calibrated XGBoost pipeline; a React + Tailwind dashboard walks operators through a multi-step customer profile wizard and returns probability, confidence, and risk verdicts.

Training is production-oriented: OneHot preprocessing, stratified CV hyperparameter search, isotonic probability calibration, and threshold optimization aimed at maximizing recall under a precision floor (business-aligned objective, not vanity accuracy).

Benchmarking across model families selected XGBoost as the winner, with optimized metrics around ~0.66 recall, ~0.60 precision, and ROC-AUC ~0.84 — packaged as deployable model artifacts with a documented report.

What it does

• Multi-step prediction wizard (demographics → services → contract → financials)
• Animated churn probability gauge and risk verdict
• Calibrated probabilities with tunable decision threshold
• Business objective: maximize recall with precision ≥ 0.60
• Feature engineering (avg monthly spend, active services count)
• Model benchmarking leaderboard across families
• Dark professional UI, fully responsive
• REST API: POST /predict and GET /health

Where it fits

• Telecom and subscription businesses ranking who is likely to leave
• Retention teams prioritizing outreach when churn risk crosses a threshold
• Data science teams demoing calibrated ML in a product UI, not only notebooks
• Product managers stress-testing threshold trade-offs (precision vs recall)

The challenges

• Imbalanced churn labels and choosing metrics that match business cost
• Calibrating probabilities so operators can trust risk scores
• Packaging training artifacts for repeatable inference in Flask

What I took away

• Thresholds should optimize business objectives, not default 0.5
• Calibration and feature engineering move the needle as much as model choice
• A clear API + wizard UX makes ML usable for non-ML stakeholders