Built a comprehensive machine learning system that identifies customers likely to cancel their telecommunications service and optimises retention campaign spending for maximum business value.

The Challenge

A telecommunications company was losing significant revenue to customer churn. They needed a system that could not only predict which customers would leave, but also optimise retention spending to maximise ROI.

Solution Approach

- Data Analysis: Analysed 7,000+ customer records to identify key churn indicators, discovering the "premium but unprotected" customer segment
- Feature Engineering: Created 30+ engineered features capturing customer behavior patterns and value indicators
- Model Development: Tested multiple algorithms from logistic regression to XGBoost; simple models surprisingly outperformed complex ones
- Business Optimisation: Built custom optimiser that adjusts prediction thresholds based on retention costs and success rates
- Production System: Developed RESTful API with comprehensive error handling, logging, and monitoring

Technical Implementation

The system uses a modular architecture with separate components for data processing, model training, prediction serving, and monitoring. All components are containerised for easy deployment and scaling.

Key Results

- Achieved 84% AUC-ROC with business-optimised thresholds
- Delivers 245% average ROI on retention campaigns
- Processes predictions in under 100ms
- Maintains 95% test coverage with comprehensive test suite

Engineering Highlights

- Implemented MLOps best practices including experiment tracking with MLflow
- Built real-time monitoring dashboard using Streamlit
- Created comprehensive logging system for production debugging
- Designed for horizontal scaling with stateless API design

The system demonstrates end-to-end ML engineering capabilities from initial data analysis through production deployment, with a focus on delivering measurable business value.