How I Built an Agentic Fraud Investigator Using TigerGraph + LangChain in 24 Hours

The Challenge

At HHGOA 2026, we were given a real-world fraud detection problem:
investigate 20 complex fraud cases using an AI agent powered by TigerGraph.

The result? A fully autonomous Fraud Investigation Agent that detects fraud rings,
retrieves compliance policy via GraphRAG, and recommends Next Best Actions — all in real-time.

Architecture

Why TigerGraph?

Standard SQL or vector databases miss relationship fraud entirely.
A money mule network is invisible in a flat table — but TigerGraph's
multi-hop traversal exposes the entire ring instantly via GSQL:

sql
SELECT src, dst FROM Account-(SHARED_DEVICE>)-Device
WHERE Device.risk_score > 0.8

5 Fraud Patterns Detected
Pattern Code    Risk
Card Testing    PT001   High
Account Takeover    PT002   High
Synthetic Identity  PT003   Medium
Money Mule Network  PT004   Critical
Merchant Collusion  PT005   High
The Agent Tools (LangChain)
The ReAct agent autonomously selects from 8 tools:


get_transaction_history
 — GSQL query on Account node

get_connected_accounts
 — multi-hop graph traversal

detect_fraud_patterns
 — GSQL algorithm execution

get_fraud_policy
 — ChromaDB GraphRAG retrieval

manage_case
 — write findings back to graph

execute_action
 — block/freeze/escalate
GraphRAG = No Hallucinated Actions
Our agent never invents a recommendation. All actions are grounded in real bank policy retrieved from a ChromaDB vector store built from our 

fraud_policy.txt
 document. This is GraphRAG in practice.

Results
✅ All 20 benchmark cases investigated autonomously
✅ SAR (Suspicious Activity Report) generated per case
✅ NBA (Next Best Action) returned for each case
✅ Full live UI at https://tigergraph-2cd6jdrgrem2jblxb49zjj.streamlit.app/
GitHub
https://github.com/foxmaster77/tigergraph

Built solo in under 24 hours for #HHGOA2026 🚀