Skip to content
Finance

Real-Time Fraud Scoring

Streaming transaction scoring with reason codes that analysts can read, for a digital lender.

p99 latency
<80ms
Fewer false positives
31%
Stack
KafkaXGBoostPythonRedis
The challenge

Rule-based checks were blocking good customers while missing new fraud patterns.

Our approach
  1. 01Engineered behavioural features from event streams
  2. 02Trained gradient-boosted models with explainability output
  3. 03Shadow-deployed for four weeks before switching traffic
Start here

Tell us what you want to change.

Bring the problem, the number it should move and your deadline. We will come back with a plan within a week.