Finomaly

Accomplishments

• Engineered a real-time fraud detection pipeline processing 6.3M+ transactions using a PyTorch Autoencoder and GraphSAGE GNN, achieving 0.951 ROC-AUC, 99.76% recall, and 4.8ms per-transaction scoring latency.
• Built streaming inference with Redpanda and Redis-cached GNN embeddings, delivering end-to-end detection in <10ms p95 and publishing flagged transactions to downstream alert consumers.
• Developed a FastAPI inference API with model explainability and validated complementary model contributions through an ablation study, surfacing human-readable signals including reconstruction error, balance inconsistencies, and fraud-ring proximity.