Agentic Fraud Investigation with TigerGraph
Agentic Fraud Investigation with TigerGraph Introduction Fraud investigation becomes difficult when transaction data, customer information, and related entities are spread across different records. For the TigerGraph × HHGoa 2026 challenge, we built an Agentic Fraud Investigation solution to analyze fraud cases and generate structured investigation results. Our solution uses TigerGraph Savanna as…
Fraud investigation becomes challenging when transaction data, customer information, and related entities are scattered across various records. To address this issue, the TigerGraph × HHGoa 2026 challenge saw the development of an Agentic Fraud Investigation solution. This solution utilizes TigerGraph Savanna as the graph layer and a custom Python-based investigation agent to analyze fraud cases and generate structured investigation results. The project processed 20 benchmark cases, ranging from HHG-001 to HHG-020.
Key components of the solution include:
1. TigerGraph Savanna as the graph layer, which stores Transaction, Customer, and Card entities and their relationships.
2. A custom Python investigation agent that reads prepared investigation data, applies fraud detection rules, calculates a heuristic risk probability, and generates investigation decisions.
3. The agent employs rule-based fraud pattern detection and heuristic risk scoring but does not utilize an external LLM API.
The workflow of the solution is as follows:
- Case Input
- Data Preparation
- TigerGraph Graph (SavannaTransactionGraph)
- Transaction and Entity Relationships
- Python Investigation Agent
- Fraud Pattern Analysis
- Risk Score
- Investigation Decision
- JSON Case Output
The graph representation facilitates the organization of relationships between transactions and related entities for fraud investigation purposes. The final results are stored as individual JSON case files, containing structured investigation information for each case, such as the detected patterns and investigation decision.
The project's GitHub repository, located at https://github.com/Bhagyashrikale01/agentic-fraud-investigation-tigergraph, contains the complete source code and case outputs. However, the current implementation has some limitations, including the use of heuristic risk probabilities, reliance on preprocessed evidence, and the absence of live external integrations.
Future improvements could involve implementing full TigerGraph MCP integration, incorporating GraphRAG-based investigation, utilizing LLM-assisted investigation reasoning, integrating live data, employing more advanced fraud detection models, and enhancing explainability and investigation reports.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
This story
This is one outlet's version. Read the fullest account.
