Design an end-to-end machine learning system for real-time payment fraud detection that processes 10k+ QPS with sub-100ms latency, specifically addressing the challenges of extreme class imbalance, feature freshness (velocity features), and delayed labels from chargebacks.