Hybrid AI-Graph System Separates LLM Reasoning from Fraud Policy Decisions

Engineers at the TigerGraph × Hacker House Goa IEEE Fraud Challenge built an autonomous fraud detection system that prevents Large Language Models from directly controlling banking actions such as blocking cards or filing regulatory reports. The architecture splits responsibilities across three tiers: an LLM reasoning layer, a TigerGraph topology layer processing over 590,000 transactions, and a deterministic Python-based policy engine that makes all final decisions. The design addresses key risks of unconstrained LLMs in finance, including hallucinations, non-reproducible outputs, and inability to meet FinCEN and SEC audit requirements. A calibrated machine learning model achieving an AUC of 0.9950 feeds into a rules-based policy engine, ensuring decisions are mathematically deterministic and auditable. The team argues this separation is essential in production fraud defense, where false positives from raw ML scores routinely disrupt legitimate customers and generate regulatory liability.