Building FraudAgent: Autonomous Financial Crime Intelligence with TigerGraph & LangGraph

VIJENDER YADAV — Thu, 24 Sep 2026 18:27:37 +0000

Autonomous Financial Crime Intelligence: Building an Agentic Fraud Investigation System with TigerGraph & LangGraph

Submission for the TigerGraph Agentic Fraud Investigation Hackathon (HHGOA)

Demo Video: Watch 3-5 Min Walkthrough

GitHub Repository: https://github.com/desmond009/Fraud_agent_HHG

Benchmark Dataset: IEEE-CIS Financial Crime Dataset (590k+ Transactions, 13.5k Entities, 5.5k Historical Closed Cases)

Evaluation: 20 / 20 Benchmark Cases Validated (100% Policy Schema Compliance)

1. What We Built

Financial crime investigation teams at tier-1 banks are under relentless pressure. Traditional rule engines and isolated machine learning models flag thousands of alerts daily, creating severe fatigue:

Fragmented Context: Analysts must manually query disparate databases (transaction ledgers, device profiles, customer profiles, regulatory blacklists).
Uncertain Signals: A transaction with an initial risk score of 0.65 may be an innocent cardholder traveling abroad, or the opening move of a distributed card-testing ring.
Slow Response Latency: Gathering evidence across multiple systems often takes days, by which time funds have already been laundered.

To solve this, we built FraudAgent — an autonomous, policy-governed fraud investigation platform that combines TigerGraph, TigerGraph MCP, LangGraph, ChromaDB GraphRAG, and a Palantir/Linear-inspired React 19 Workspace.

Instead of treating fraud detection as a static point prediction, FraudAgent operates as a stateful, iterative investigation agent that:

Ingests Alerts: Automatically triggers from risk score anomalies, customer dispute reports, or analyst escalations.
Traverses Knowledge Graphs: Queries multi-hop relationship linkages across customers, cards, devices, and historical fraud rings using TigerGraph GSQL algorithms.
Calibrates Uncertainty: Evaluates fraud probability before and after evidence collection, computing explicit explanations of what changed.
Enforces Policy Governance: Maps findings against bank regulations (Rules R1–R10) with role-based sign-offs (auto, L1 Lead, L2 Manager).
Generates Regulatory Deliverables: Instantly drafts official FinCEN Suspicious Activity Reports (SARs) with full audit trails.
Writes Back to Case Memory: Persists investigation outcomes back into the graph, continuously training and expanding institutional memory.

2. System Architecture

FraudAgent is designed around an event-driven, microservices-ready architecture:

┌────────────────────────────────────────────────────────────────────────────────┐
│                           REACT 19 ANALYST WORKSPACE                           │
│  - Spring Force Graph Canvas      - Uncertainty Delta Visualizer               │
│  - Case Progression Timeline      - L1 / L2 Approval Sign-off Actions          │
│  - Interactive Evidence Dossier   - FinCEN SAR Regulatory Drawer               │
└───────────────────────────────────────┬────────────────────────────────────────┘
                                        │ REST / JSON (FastAPI Bridge)
                                        ▼
┌────────────────────────────────────────────────────────────────────────────────┐
│                       FASTAPI BACKEND & ORCHESTRATION                          │
│  - /api/cases, /api/graph, /api/audit-log, /api/model/predict, /api/model/train│
└──────────────────┬────────────────────┬────────────────────┬───────────────────┘
                   │                    │                    │
                   ▼                    ▼                    ▼
     ┌──────────────────────┐ ┌───────────────────┐ ┌──────────────────────┐
     │ 6-NODE LANGGRAPH     │ │ CHROMA VECTOR RAG │ │ ML INFERENCE PIPELINE│
     │ INVESTIGATION STATE  │ │ Policy Rules R1-10│ │ HistGradientBoosting │
     │ MACHINE              │ │ & FinCEN Typologies││ Checkpoint Persistent│
     └──────────┬───────────┘ └───────────────────┘ └──────────────────────┘
                │ Tool Calls via Model Context Protocol (MCP)
                ▼
┌────────────────────────────────────────────────────────────────────────────────┐
│                            TIGERGRAPH CLOUD / MCP                              │
│  • Graph Schema: Customer, Card, Transaction, DeviceProfile, ClosedCase        │
│  • GSQL Graph Algorithms: Multi-hop Ring Traversal, Temporal Velocity Windows  │
│  • Vector Graph Memory: Writeback of resolved cases for historical retrieval   │
└────────────────────────────────────────────────────────────────────────────────┘

The 6-Node LangGraph State Machine

trigger_ingestion: Normalizes alert metadata (risk scores, customer reports, transaction details).
initial_investigation: Dispatches TigerGraph MCP queries to inspect account history, transaction windows, and multi-hop entity neighborhoods.
evidence_synthesis: Synthesizes graph evidence against bank policies retrieved via ChromaDB GraphRAG.
uncertainty_assessment: Measures signal confidence and computes initial fraud risk and initial Next-Best Action.
evidence_simulation: Gathers controlled additional evidence (e.g. simulated customer confirmation/denial or step-up authentication), recalculating calibrated probability.
case_memory_writeback: Commits resolved case records, SAR filings, and approval decisions into TigerGraph as new ClosedCase vertices.

3. How TigerGraph is Used

TigerGraph serves as the persistent nervous system of our platform:

A. Graph Schema (`schema/schema.gsql`)

Vertices:
- Customer: Account owner metadata.
- Card: Payment cards linked to accounts.
- Transaction: Individual financial authorizations ($amount, $timestamp, risk score, channel).
- DeviceProfile: MD5-hashed digital fingerprint (DeviceInfo, OS version, browser, screen resolution).
- EmailDomain & BillingRegion: Contextual anchor points.
- ClosedCase: Prior investigations serving as dynamic case memory.
Edges:
- Customer -OWNS-> Card
- Card -MADE-> Transaction
- Transaction -FROM_DEVICE-> DeviceProfile
- Transaction -PURCHASER_EMAIL-> EmailDomain
- Transaction -BILLED_IN-> BillingRegion
- Transaction -NEXT-> Transaction (temporal sequence for card testing detection)
- ClosedCase -INVOLVES-> Transaction
- ClosedCase -ON_CARD-> Card
- Card -CONNECTED_TO-> Card (device-sharing fraud rings)

B. High-Performance GSQL Queries

get_card_transaction_window: Computes rolling transaction frequency and amount velocity within 24 hours of an alert.
get_device_region_neighborhood: Traverses multi-hop paths to uncover shared devices, shared billing addresses, and linked cards with prior fraud history.
match_historical_cases: Fetches similar resolved cases by fraud typology and customer behavior.
write_case_memory: Inserts a ClosedCase vertex and connects it to affected cards and transactions in real-time.

C. TigerGraph MCP (Model Context Protocol) Bridge

Rather than feeding raw data into an LLM prompt, the agent invokes structured MCP tools:

tg_get_transaction(txn_id)
tg_query_txn_window(card_id, timestamp, window_hours)
tg_query_neighborhood(txn_id)
tg_match_historical_cases(pattern, max_results)
tg_write_case_memory(case_id, verdict, pattern, exposure, ...)

4. Agentic Capabilities Implemented

1. Two-Stage Uncertainty Calibration

In high-stakes financial operations, models cannot act blindly on initial indicators. FraudAgent separates action recommendations into two distinct stages:

Initial Next-Best Action: Formulated before requesting additional evidence (e.g., VERIFY_WITH_CUSTOMER (auto), STEP_UP_AUTH (auto)).
Final Next-Best Action: Formulated after additional evidence is received (e.g., customer confirms authorized activity $\rightarrow$ CLOSE_NO_FRAUD (auto); customer denies activity $\rightarrow$ BLOCK_CARD (L1), FILE_REPORT (L2)).
What Changed Explanation: Generates plain-English rationale for why the confidence shifted (e.g., "Customer denial confirmed unauthorized usage, raising fraud probability from 54% to 88%").

2. Policy-Governed Gating (L1 / L2 Approval Workflow)

Automated AI actions are bounded by strict banking regulations:

auto: Autonomous execution (logging, monitoring, closing cleared alerts under Rule R3).
L1 (Lead Analyst): Sign-off for single card blocks ($\le \$2,500$) and authorization declines.
L2 (Fraud Manager): Sign-off for high-exposure freezes ($> \$2,500$), blocking all account cards, and external FinCEN SAR submissions.

3. Automated FinCEN Suspicious Activity Reports (SAR)

When exposure exceeds \$1,000 or organized fraud rings are detected, the agent drafts complete regulatory narratives answering Who, What, When, Where, Why, and How, ready for compliance export.

5. What We Learned

Graph Traversal Disarms Fraud Rings: Individual transactions often look completely harmless in isolation (e.g. a \$50 online purchase). However, traversing the TigerGraph neighborhood reveals that the device fingerprint has been used across 110 other cards with prior fraud convictions. Graph context is irreplaceable.
GraphRAG Beats Raw Prompting: Indexing policy rules into a vector store and retrieving relevant clauses based on agent findings ensures that actions strictly conform to bank procedures without hallucinated policies.
Approval Gating Builds Trust: Analysts reject "black-box" decisions. By splitting recommendations into explicit approval tiers (auto, L1, L2) with clear justifications, human teams remain in control while saving 90% of investigation time.

6. What We Would Improve With More Time

Real-time Streaming Graph Updates: Ingesting Kafka transaction feeds directly into TigerGraph via continuous streaming loaders.
Graph Neural Network (GNN) Embeddings: Incorporating TigerGraph Graph Convolutional Networks (GCN) to predict ring topologies before transactions even execute.
Voice-Interactive Analyst Copilot: Adding a multimodal speech interface for analysts to interact with graph subgraphs in hands-free triage rooms.
Automated Clearing House (ACH) Inter-Bank Network: Expanding graph traversals across participating institutional graphs using privacy-preserving federated graph analytics.

Built with passion for the TigerGraph Agentic Fraud Investigation Hackathon.

DEV Community: VIJENDER YADAV