Payment fraud scoring system

Design a real-time transaction fraud-detection system for a payments platform processing 30K transactions/s that must approve/decline/step-up each transaction in under 100ms. It scores account-takeover, stolen-card, and collusion fraud using historical behavior, device signals, and velocity features, and must adapt to new fraud rings within hours. Threat model: fraudsters probing thresholds, fast-moving fraud rings, and the cost asymmetry between blocking a good customer and approving a fraudulent charge. Cover the feature/serving layer, the model and rules, the feedback loop with delayed labels, and the decision policy.