Fraud Detection Rule Refresh
Rebuilt fraud pipeline for higher recall and lower latency using real-time ML
Situation
Legacy fraud detection system produced high false positives and missed emerging fraud patterns, costing the business significantly.
Task
Rebuild fraud detection pipeline to achieve higher recall, lower false positive rate, and reduce detection latency.
Action
Business & Strategy
Worked closely with finance and risk teams to define acceptable risk thresholds and business rules. Established feedback loop with fraud investigation team for continuous model improvement.
Technical Implementation
Implemented feature store architecture for real-time feature computation. Deployed XGBoost model with online learning capabilities. Optimized inference pipeline to achieve sub-100ms latency. Built monitoring dashboard for model drift detection.
Results
Business Impact
Reduced fraud losses by approximately $250K per quarter. Improved customer experience by reducing false positive blocks by 40%.
Technical Achievement
Increased recall by 6.2 percentage points at iso-precision. Achieved <100ms detection latency. Model F1 score: 0.89.