Capstone · Bluemind Solutions AI/ML Internship · 2026
Customer Churn Intelligence System
A production-ready customer churn intelligence system completed as the capstone for Bluemind Solutions' Core AI & ML Internship.
0.849
CV ROC-AUC
82%
churner recall
~$179k
est. annual recovery
The problem
Customer churn predictions are only useful when the evaluation is leak-free, the threshold reflects business cost, and the result can be used by someone beyond a notebook.
Constraints
- Build a leak-free pipeline on the IBM Telco Customer Churn dataset.
- Compare multiple classical and gradient-boosted models with stratified 5-fold cross-validation.
- Turn model output into an actionable workflow through a live dashboard for single and batch scoring.
What I built
- A preprocessing pipeline using Scikit-learn Pipelines and ColumnTransformer to keep transformations inside validation.
- Model comparison across Logistic Regression, Random Forest, XGBoost and LightGBM.
- Business-driven threshold tuning, error analysis and SHAP-style analysis for more useful decisions than a default 0.5 cutoff.
- A Streamlit dashboard for single-customer and batch churn scoring.
Trade-offs
LightGBM as the final model
It delivered the strongest reported result in the comparison: 0.849 cross-validation ROC-AUC with 82% recall on churners.
Threshold tuning instead of optimising accuracy alone
The system is meant to support retention decisions, so missing a likely churner matters more than maximising a generic accuracy score.
Streamlit dashboard alongside the model pipeline
A usable scoring surface makes the capstone demonstrable as a shipped system, not only as an experiment.
Outcome
- LightGBM reached 0.849 CV ROC-AUC and 82% recall on churners.
- Estimated ~$179k in annual recoverable revenue from the business-oriented scoring setup.
- Completed as a rigorous 4-week (~120 hour) Core AI & ML Internship capstone at Bluemind Solutions Pvt. Ltd.
Stack
PythonNumPyPandasScikit-learnXGBoostLightGBMStreamlit
Want the long version?
Happy to walk through the decisions live.