Feature store
User's transaction history, velocity (txns/hour), device fingerprints. Real-time from Redis/Aerospike. Historic from data warehouse.
Feature store
User's transaction history, velocity (txns/hour), device fingerprints. Real-time from Redis/Aerospike. Historic from data warehouse.
Rules engine
Explainable, precise, fast. 'Card used in country A + country B within 1 hour → decline'. Business writes rules; engineers wire them.
ML model
Gradient boosted trees or shallow neural net. Runs in <10ms. Scores based on 100+ features. Explainability harder than rules.
Graph analysis
Same device + different cards → shared identity. Ring of accounts moving money → likely money laundering. Real-time graph queries.
Feedback loop
Confirmed frauds + false positives labeled. Retraining pipeline updates models weekly. Rules updated by risk team continuously.