HACKATHON DEMO · REAL-TIME
Multi-Agent
Fraud Detection
A hierarchical AI system where Claude acts as the brain —
orchestrating specialized agents to catch fraud in real time
Claude 3.5 Sonnet · GraphSAGE · Redis Stream · MCP Tools
The Problem
Fraud hides in connections,
not individual transactions
Money laundering works by moving funds through a ring of accounts. Each transfer looks normal — the fraud only appears when you look at the whole network.
Real-time
OUR DETECTION SPEED
Architecture
4 Agents, 1 Brain
📡
Data Agent
Redis Stream · live feed
🕸️
Graph Agent
NetworkX · PyG · topology
⟷
🧠
Orchestrator
Claude 3.5 Sonnet
ReAct Loop
⟷
🧬
Inference Agent
GraphSAGE · /predict
🚨
Alert Agent
Structured output · dispatch
Step 2 — Standardized Interfaces
MCP Tools: clear contracts
between Claude and each agent
Step 3 — Agentic Loop
ReAct: Reason → Act → Observe
Claude decides how deep to dig
💬
Transaction In
New suspicious tx
→
🤔
Reason
Claude picks tool
→
→
→
📋
Report
Enough evidence?
↩ If more evidence needed — loop back and call another tool autonomously
Claude autonomously decides: score first → check history → verify graph topology → dispatch
Step 4 — Automated Response
3-Level Risk Dispatcher
🚫 BLOCK + Full Trace Report
Transaction intercepted · Orchestrator triggers multi-hop chain tracing · Structured report generated
👁️ HUMAN REVIEW Queue
Transaction allowed · Flagged for analyst review · Human-in-the-loop decision
✅ ALLOW + Async Update
Fast pass-through · Graph features updated asynchronously in background
Step 5 — Hackathon Demo
Catching a money laundering ring
1
A $181,097 TRANSFER comes in — looks like a normal transaction on the surface
2
Claude calls predict_fraud → GNN scores sender at 0.85 — high risk flag triggered
3
Claude calls get_account_history → 14 transactions in 24 hours, classic structuring pattern
4
Claude calls get_graph_topology → detects 4-node cycle — circular fund flow confirmed
5
Claude calls dispatch_action(HIGH) → transaction BLOCKED + full investigation report generated
Summary
Modular, decoupled,
and production-ready
Model-agnostic Orchestrator — swap GraphSAGE for any GNN without changing Claude's logic
Standardized MCP Tools — clear JSON Schema contracts between every agent
ReAct Loop — Claude reasons autonomously, calls tools based on evidence, not fixed rules
3-Level Dispatcher — automated response with human-in-the-loop for borderline cases
LIVE DEMO
python -m orchestrator.demo_runner