Fraud Detection — Kafka Streaming Pipeline

A re-engineered credit card fraud detection pipeline achieving 10-100x performance improvement over the original proof-of-concept. Multi-threaded consumer and predictor workers allow independent scaling. Batch processing (100–500 transactions per batch) enables highly efficient ML model inference. Average latency reduced from 100ms+ down to 10–30ms. Built-in observability reports throughput, queue sizes, and latency in real time. Horizontally scalable via Kafka consumer groups. Model metrics: Precision 0.928, Recall 0.786, F1 0.851, ROC-AUC 0.944.