Design a Fraud Detection System for Fintech

Design a real-time fraud detection system for a payment processing platform that handles millions of transactions daily. The system must score every transaction for fraud risk within strict latency limits and decide whether to approve, block, or flag for manual review.

Scale Requirements

• 5 million transactions per day
• Decision latency: under 200ms end-to-end
• Fraud rate: approximately 0.1% of transactions (highly imbalanced)
• Cost of false positive (blocking legitimate transaction): $50-200 in lost revenue and customer churn
• Cost of false negative (missing fraud): $500-5000 average per incident

Design Requirements

1. Design the real-time scoring pipeline from transaction event to decision.
2. Explain your approach to handling extreme class imbalance.
3. Design the decision system that balances automated blocking vs. manual review.
4. Address adversarial robustness — fraudsters adapt to your model.
5. Describe your monitoring and incident response strategy.