An Agent That Knows When It Doesn't Know: Fraud Investigation on TigerGraph

Developers at the TigerGraph x Hacker House Goa hackathon have created an agentic fraud investigation system designed to improve detection accuracy through graph-based reasoning. By utilizing the IEEE-CIS dataset, the system employs a LangGraph agent integrated with TigerGraph to perform complex traversals and vector searches. Unlike traditional models that rely on static risk scores, this agent uses Bayesian updates to calibrate its confidence levels. When uncertainty arises, the system proactively requests additional evidence, such as customer verification or analyst review, before making a final decision. The architecture leverages GSQL queries, graph algorithms like Louvain and PageRank, and vector search to identify coordinated fraud patterns that standard models often miss. This approach demonstrates the effectiveness of combining structured graph data with LLM-driven reasoning, allowing for more transparent, audit-ready, and adaptive fraud prevention workflows in financial services.