Building TRACE: Graph-Based Fraud Investigation with TigerGraph and Gemini

The TRACE project, developed for the TigerGraph Agentic Fraud Investigation challenge, introduces a prototype system designed to enhance fraud detection through graph-based analysis. By integrating TigerGraph for relationship mapping and vector retrieval with Gemini 2.5 Flash for evidence synthesis, the system provides explainable investigation records. TRACE utilizes a local embedding model to overcome API limitations, allowing for semantic search across transaction data, historical cases, and policy documents. The architecture focuses on connecting complex entities—such as device profiles, billing regions, and customer history—to identify suspicious patterns while maintaining transparency regarding uncertainty. Although currently a prototype that simulates customer responses and provides policy-based recommendations rather than executing banking actions, TRACE demonstrates a robust framework for evidence-driven investigations. Future iterations aim to incorporate adaptive tool selection and more precise document retrieval to further improve the accuracy and explainability of fraud assessments.