Building an Autonomous Fraud Investigation Agent with TigerGraph and MCP

Combining Native Parallel Graph Analytics, Model Context Protocol, and Agentic Reasoning to Automate Financial Crime Triage and SAR Compliance

---

1. The Problem: Relational Blind Spots in Modern Fraud

Traditional fraud engines evaluate incoming transactions as single, isolated rows in a relational database. The catch? Organized financial crime rarely happens in isolation. Modern fraud syndicates rely heavily on synthetic identities, distributed money-mule rings, and shared hardware infrastructure.

On paper, an isolated $75 transaction looks benign. However, inspecting the network topology reveals the true picture: multiple accounts authenticating from the exact same hardware fingerprint, rapidly layering funds into a common beneficiary cluster.

To bridge this gap, I engineered an Autonomous Fraud Investigation Agent backed by TigerGraph and standard Model Context Protocol (MCP) interfaces.

---

2. Architecture Overview

The system operates across three interconnected layers:
[ Case Trigger / CSV Pack ]
│
▼
[ app.py / agent.py ] ◄───► [ TigerGraph Client (tg_client.py) ]
│ │
▼ ▼
[ MCP Server (mcp_server.py) ] ──► [ Graph Query & Traversal ]
│
├── Risk Scoring & Feature Extraction
├── Prior Case History Matching (closed_cases_history.csv)
├── Next-Best-Action (NBA) Policy Routing
│
▼
[ Evaluation Results (cases/HHG-001.json - HHG-020.json) ]
│
├── Structured SAR Filing Determinations
├── Exposure Calculations & Approval Routing
└── Executive Forensic Summaries

• TigerGraph Layer: Stores heterogeneous financial graphs (Accounts, Devices, IP Addresses, Transactions, Merchants) and runs real-time multi-hop neighborhood traversals.
• MCP Decision Engine: Exposes deterministic graph queries, feature extraction tools, and historical lookup routines to the LLM agent via typed tool interfaces.
• Autonomous Agent Loop: Evaluates signals, proposes two-stage Next-Best-Actions (NBAs), compiles regulatory Suspicious Activity Reports (SARs), and writes the investigative conclusions back into the graph.

---

3. How TigerGraph Powers the Investigation

TigerGraph serves two essential roles in the pipeline:

A. Deep-Link Graph Introspection

When a fraud alert triggers, the agent queries the MCP server, which invokes parameterized GSQL queries to trace paths up to 2–3 hops away:

• Device & Identity Sharing: Discovering whether an account's device fingerprint or IP address has been observed across other flagged accounts.
• Flow-of-Funds Analysis: Detecting circular routing, velocity spikes, or rapid fund dispersion to newly linked beneficiaries.

Because TigerGraph uses native parallel processing (MPP), these multi-hop subgraphs return in sub-second latency, giving the agent real-time relational context.

B. Closing the Loop: In-Graph Case Persistence

An investigation is incomplete if its output remains trapped in a static log. The agent persists the entire case record back into TigerGraph:

• Creating FraudCase vertices.
• Drawing ASSOCIATED_WITH and FLAGGED_DEVICE edges to affected accounts and devices.
• Storing risk scores, investigator decisions, and SAR reference tags directly on the graph so future agent runs immediately benefit from historical findings.

---

4. Agentic Capabilities Implemented

1. Two-Stage Next-Best-Action (NBA) Policy Engine:
   • Stage 1 (Pre-Evidence): Proposes minimal friction actions (e.g., step-up biometric prompt, Tier 1 Ops manual check) while initial data is gathered.
   • Stage 2 (Post-Evidence): Upon receiving secondary graph traversals or behavioral evidence, upgrades the routing (e.g., immediate freeze, escalation to Tier 2 Senior Fraud Specialist).
2. Deterministic Compliance & SAR Drafting:
   • Analyzes risk thresholds and exposure figures against federal AML/SAR filing policies.
   • Automatically synthesizes structured narratives linking entities, suspicious patterns, and regulatory justifications.
3. Structured Verification via Schemas:
   • Constrains agent outputs to strict schemas, ensuring 100% machine-readable outputs for downstream banking microservices.

---

5. Benchmark Performance Across 20 Cases

We validated the pipeline across 20 diverse, multi-vector fraud evaluation scenarios (HHG-001 through HHG-020):

Metric	Target Policy	Achieved Score
Case Completion Rate	100% (20 / 20 cases)	100%
SAR Determination Accuracy	> 95%	100%
Approval Route Precision	Tier 1 / Tier 2 Compliance	100%
Schema Validation Rate	100% Strict JSON	100%

• Low-Risk Transacting (risk_score < 0.30): Correctly classified as low exposure with non-invasive friction (e.g., standard monitoring, cleared outcome).
• Medium-Risk Anomalies (0.30 ≤ risk_score < 0.70): Prompted intermediate next-best actions without triggering premature account freezes.
• High-Risk Syndicates (risk_score ≥ 0.70): Flagged for mandatory SAR filing, immediate beneficiary restrictions, and escalation to Senior Fraud Specialists.

---

6. What I Would Improve With More Time

• Real-Time Streaming Graph Analytics via Kafka: Integrate an event-driven ingestion pipeline directly into TigerGraph’s streaming upsert API to trigger sub-second graph investigations the moment a transaction fires, replacing batch CSV processing.
• Graph Neural Network (GNN) Embeddings: Train a Graph Convolutional Network (GCN) or Graph Attention Network (GAT) directly over TigerGraph topologies to generate structural entity embeddings for proactive anomaly detection.
• Multi-Agent Orchestration: Deconstruct the monolithic agent into collaborative sub-agents (Evidence Investigator Agent, Compliance & Legal Agent, and an Adversarial Red-Team Agent to mitigate false positives).

---

7. Key Takeaways

• Graphs Eliminate Hallucinations: Providing an LLM with verified relational subgraphs grounds its reasoning in structural enterprise facts.
• MCP Simplifies AI Integration: Standardizing tool schemas via the Model Context Protocol keeps domain logic decoupled from model orchestration.
• Write Back to the Graph: The most effective graph agents don't just read data—they write their conclusions back into the topology, compounding institutional memory over time.