AI Fraud Detection Agent Uses Graph Database to Investigate 20 Real Cases

A fraud investigation agent built for a hackathon uses TigerGraph alongside Google's Gemini model to analyze transaction data, device activity, and cardholder behavior through graph traversal rather than simple prompts. The system checks flagged transactions against ten coded bank policy rules and five documented fraud patterns, assigning a fraud probability before recommending actions with defined approval routes. When evidence is insufficient, the agent requests additional input — such as customer validation or analyst review — before reassessing its decision. All case findings, including suspicious activity reports and before-and-after recommendations, are written back into the graph, forming a persistent memory for future investigations. The agent was tested against 20 benchmark cases and results are viewable through a live Streamlit dashboard without any setup required.