The challenge
Rules-only fraud screening produced a flood of false positives. Analysts spent most of the day clearing alerts that turned out to be genuine customers.
What we did
- Built a streaming pipeline that joins card transactions with device and behaviour signals
- Trained and validated scoring models alongside the existing rules, then tuned thresholds with the fraud team
- Delivered an analyst console that shows why a transaction was flagged
The outcome
Analysts work a shorter, better-ranked queue, and the bank can explain each decision to auditors.
