Design a high-throughput, low-latency machine learning system to detect and prevent fraudulent transactions in real-time. The system must handle extreme class imbalance, evolving adversarial patterns, and delayed feedback from financial institutions while minimizing impact on legitimate user experience.