How I Built a Real-Time Fraud Detection System That Handles 71,000 RPS at p95 <6ms

Sentinel is a Go-based real-time fraud detection service using an XGBoost model exported to ONNX for fast production inference.

The pipeline adds an LRU cache and Prometheus metrics to keep requests efficient and observable at scale.

Built on a highly imbalanced credit card fraud dataset, the model aims to balance accuracy with low-latency payment screening.

The system reportedly handled 7.8 million requests with zero errors and peaked at 71,000 RPS with p95 latency under 6 ms.