⚡ Fraud Graph Demo Neo4j + GNN + LLM
📦 GitHub Repo
🛡️ Graph Data Science & Anti-Money Laundering (AML) Intelligence

End-to-End Fraud Detection
Knowledge Graph & AI Agent

A production-calibrated graph intelligence workspace integrating Neo4j 5 Community + GDS 2.13, PyTorch Geometric GraphSAGE, and a LangGraph self-healing Text-to-Cypher agent to uncover sophisticated money mule rings and transaction anomalies.

Neo4j 5 GDS PyTorch GraphSAGE LangGraph Agent Louvain Community PageRank Influence Weakly Connected Components Betweenness Centrality PaySim Dataset (50k Txns)
Forensic Scenarios:
Flagged Fraud / Drain Node
Mule Relay Account
Origin / Legit Account
Cash-Out Exit Node
Forensic Node Inspector HIGH RISK
GraphSAGE GNN Risk (`fraudProb`) 0.942
PageRank
0.874
Louvain Comm
#14 (Ring)
Betweenness
0.681
WCC Component
#3 (Isolated)
Rule Engine Violation Status
⚠️ flagMule: Detected on A→B→C Layering Chain
⚠️ flagDrain: 98.4% Balance Emptied
✓ flagVelocity: Within Normal Threshold
💡 Click any node on the graph to inspect forensic telemetry.
Algorithm Engine Controls

Live Parameter & Threshold Playground

Test how adjusting business risk thresholds and GNN cutoffs from app/config.py alters classification sensitivity across the transaction topology.

GNN Fraud Threshold (`gnn_fraud_threshold`) 0.50
Cutoff score above which GraphSAGE classifies an account as fraudulent.
Account Drain % (`drain_pct`) 95%
Minimum balance percentage emptied in a single transfer to trigger smash-and-grab flag.
Velocity Txn Threshold (`velocity_tx_threshold`) > 3 txn
Number of rapid transfers within time window signaling automated card testing.
Natural Language AI Agent

LangGraph Self-Healing Text-to-Cypher Assistant

Experience how the LangChain + spaCy + DeepSeek agent converts complex forensic questions into verified Cypher queries with auto-retry self-healing.

FraudGraph Agent (LangGraph + DeepSeek-v4-flash)
Schema Verified · In-Memory GDS
Try Prompt:
FRAUD GRAPH AGENT
Hello! I am your Anti-Money Laundering (AML) Graph Investigator. You can ask me natural language questions about transaction topologies, money mule chains, GDS metrics, or GraphSAGE fraud scores.
Performance & Scale Telemetry

GDS Algorithm Optimization Benchmarks

Key performance milestones extracted from benchmark_report.md across 50,000 transaction graph scales.

Sampled Betweenness Centrality
142× Faster

Sampling budget (size=100) reduced execution time from 4,534 ms to 32 ms with negligible ranking distortion for money-relay detection.

PageRank Convergence
2 Iterations

On transaction bipartite graphs, influence scores converge in exactly 2 iterations. Budget capped at 5 to eliminate wasted compute cycles.

WCC Streaming Detection
< 20 ms

Weakly Connected Components runs in O(n+e), isolating self-contained fraudulent subgraphs in under 20 ms for real-time alerting.