{
  "id": 9363956,
  "title": "Building Casework: An Agentic Fraud Investigator with TigerGraph and GraphRAG",
  "url": "https://urgent.news/2026/09/23/building-casework-an-agentic-fraud-investigator-with-tigergraph-and",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-23T15:35:21.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/dhruv_ghosal_019265a175ed/building-casework-an-agentic-fraud-investigator-with-tigergraph-and-graphrag-nnl"
  },
  "original_language": "en",
  "account": "A suspicious transaction initiates a fraud investigation. The challenge lies in determining which evidence to collect, what remains unclear, and which actions are permissible. The Casework application, developed for the TigerGraph × HHGoa challenge, addresses this issue. Casework integrates graph evidence, document retrieval, a local language model, and policy rules to guide a fraud investigation from its initial trigger to a final recommendation. The goal is to provide a clear record of the investigation for review by analysts.\n\nCasework offers a browser interface for forensic analysis of fraud cases. Each investigation begins with a customer, a card, and a transaction that has been flagged. The software retrieves pertinent relationships from TigerGraph, evaluates transaction signals, conducts historical case searches and policy document analyses, and generates a structured case record. The interface displays the current case status, verdict, and summary; graphical evidence illustrating relationships between the transaction and related entities; a list of uncertainties, missing information, and unresolved questions; suggested actions before and after new evidence is obtained; approval pathways for recommendations requiring human input; and draft reports of suspicious activity where necessary.\n\nThe application encompasses the 20 benchmark answer files in the root cases/ folder, named HHG-001.json through HHG-020.json. Each answer includes evidence, findings, affected transactions, potential financial exposure, requests for additional evidence, and both initial and final recommended actions.\n\nBehind the scenes, an original Python agent governs a bounded investigation workflow. The primary responsibilities are as follows:\n\n- TigerGraph Savanna manages graph entities, relationships, document vectors, and case records.\n- GSQL extracts transaction context and connected evidence.\n- TigerGraph's MCP (Managed Cloud Platform) facilitates graph query capabilities within the workflow.\n- Ollama + Llama 3.2 3B generates semantic search queries, proposes evidence requests, and reviews supplied evidence.\n- nomic-embed-text creates embeddings for semantic retrieval tasks.\n- Python analysis and policy modules compute transaction signals, analyze graph neighborhood structures, assign action routes, calculate evidence relevance, and adhere to policy rules.\n- FastAPI and a browser UI present case details and allow users to initiate investigations.\n\nThe workflow progresses as follows: Investigation trigger → Graph evidence via TigerGraph MCP → Transaction signal analysis and relationship examination → Model suggests retrieval and evidence requests → Vector retrieval of policies, historical data, and learned case knowledge → Evidence review and uncertainty assessment → Policy rules and approval routing → Final validated case record and draft report → Persistence of graph data and versioned case history.\n\nA critical decision in the dataset involves the absence of card IDs for all transactions. To maintain accurate connections between transactions and cards, the application relies on historical anchors and the flagged transaction to establish relationships. This approach ensures that card-to-transaction associations are reliable for subsequent analytical conclusions.",
  "summary": "A suspicious transaction starts an investigation. The harder part is deciding what evidence to gather, what remains uncertain, and which action is justified. We built Casework for the TigerGraph × HHGoa challenge to explore that problem. Casework is a fraud investigation application that connects graph evidence, document retrieval, a local language model, and policy rules. It follows a case from…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}