Fraud scoring with adaptive features

Design a real-time fraud-detection pipeline for a payments processor handling 30k transactions/sec that must score each transaction (approve/decline/challenge) within ~100ms inline with checkout. It needs per-entity historical features (card velocity, device history, merchant risk), must adapt to new fraud patterns quickly, and must keep false declines low because they cost legitimate revenue. Cover the scoring path, the feature store, how you balance latency against feature richness, and how you update models against an adversarial attacker.