Autonomous Fraud Investigation Agent: Scaling Graph Intelligence with TigerGraph & LangGraph

1. The Real-World Fraud Problem Fraud teams in modern financial institutions face an overwhelming asymmetry: transaction volumes surge by millions daily, while manual fraud investigation remains slow, fragmented, and resource-constrained.

Traditional rule engines and machine learning models output isolated risk scores. However, a score alone is not an answer:

A raw risk score cannot explain why an alert fired.
It cannot evaluate whether the evidence is sufficient or conflicting.
It cannot determine the cheapest safe action that balances fraud loss against customer insult and churn.
For the TigerGraph Hacker House Goa '26 Challenge, we engineered an end-to-end, agentic fraud investigation system that transforms uncertain graph signals into defensible, cost-weighted decisions with zero hallucinations.

1. System Architecture Our solution combines TigerGraph as the high-performance distributed graph traversal engine and LangGraph as the deterministic, stateful agent orchestrator:

[ Incoming Alert / Transaction ID ]
│
▼
┌─────────────────────────────────┐
│ TigerGraph Graph Engine │ <-- Multi-hop queries (Cards, Devices,
└────────────────┬────────────────┘ Email Hubs, Coarse Address Codes)
│ Structured Subgraph Evidence
▼
┌─────────────────────────────────┐
│ Uncertainty & Conflict Engine │ <-- Evaluates evidence completeness,
└────────────────┬────────────────┘ signal quality, and data conflicts
│
▼
┌─────────────────────────────────┐
│ GraphRAG Context Builder │ <-- Assembles grounded historical cases
└────────────────┬────────────────┘ and recurring pattern memory
│
▼
┌─────────────────────────────────┐
│ LangGraph Agent Orchestrator │ <-- Versioned prompt & state loop:
└────────────────┬────────────────┘ Investigate -> Assess -> Decide
│
▼
┌─────────────────────────────────┐
│ Economic Cost-Benefit Engine │ <-- Weighs Exposure ($) against
└────────────────┬────────────────┘ Customer Friction Costs ($)
│
▼
[ Traceable Next-Best Action + Claim-Level Provenance Audit Trail ]

1. How We Leveraged TigerGraph TigerGraph provides sub-second multi-hop graph traversals that allow our agent to look beyond single transactions:

Transaction Context (get_transaction_context): Extracts seed attributes (amount, timestamp, billing codes) and links to primary payment entities.
Shared Card Rings (find_shared_card_activity): Traverses across card fingerprint vertices to find related transactions and accounts using identical payment instruments.
Device Farms (find_shared_device_activity): Connects digital fingerprints and user-agent strings to detect coordinated botnets and emulator farms.
Shared Address & Email Hubs (find_shared_address_activity, find_shared_email_activity): Identifies convergence points across billing region codes and high-velocity email domains.
2-Hop Network Neighborhoods (investigate_transaction_network): Performs a 2-hop topological fan-out across the graph to assess systemic contagion and ring size.

1. Key Differentiators & Breakthroughs Differentiator 1: Cost-Based Economic Decisions A $15 charge and a $2,500 charge with identical risk signals should not receive the same treatment. Blocking a $15 purchase creates up to $60 in customer insult and dispute costs to stop a $10 risk exposure.

We built an Economic Cost-Benefit Arbitration model:

Net Economic Benefit

Expected Fraud Loss
−
Customer Friction Cost

Micro-transactions ($\le
20
): Punitive blocks are downgraded to customer warnings (WARN_CUSTOMER), saving operational overhead while preserving customer trust.
High-value charges ($\ge
1
,
000
): Monitoring actions are escalated to full case investigations (CREATE_CASE) with mandatory SENIOR_ANALYST approval routes.
Differentiator 2: Discovered Undocumented Pattern Mining
Standard fraud systems only check documented rules. By analyzing graph topological signatures of confirmed fraud cases versus cleared cases from the benchmark closed cases, our agent discovered an undocumented fraud topology: "Distributed Identity Hub / Synthetic Mule Ring"

Topological Signature: Transactions converging on shared cards or devices while rotating across disparate email addresses.
Empirical Ground Truth: Present in 78.6% of confirmed fraud cases vs only 4.2% of cleared cases—an empirical 18.7x relative risk / odds ratio.
Differentiator 3: Claim-Level Provenance (Zero Hallucinations)
LLMs hallucinate when they synthesize unstructured conclusions. Our orchestrator enforces strict claim-level provenance tokens on every finding: [ | :]

Example:
[find_shared_card_activity | shared_card:18227|583.0|150.0|226.0] 12 other transaction(s) share the same card identifier.

Every finding directly links back to an immutable GSQL graph query and vertex ID.

1. Benchmark Results & Verification 20 / 20 Official Benchmark Cases: Successfully evaluated and exported to cases/HHG-001.json through cases/HHG-020.json matching official HHGOA specifications. Test Suite: Over 370 automated unit and API tests passing with 100% test success rate. Production UI: Modern Next.js dashboard providing interactive graph inspection, uncertainty breakdowns, and audit trails.
2. Conclusion By pairing TigerGraph's graph traversal capabilities with LangGraph's deterministic agent loop and cost-benefit arbitration, we demonstrated that agentic AI can move from experimental chat interfaces into mission-critical, auditable, and economically sound financial defense.

Explore the codebase and test cases on GitHub:
👉 https://github.com/HemanthSai-G/HHGOA_26