LLM Agents Debate Bitcoin Fraud

I built a real-time Bitcoin fraud detection system where three LLM agents debate whether a transaction is fraudulent:

• Agent 1 (Prosecutor): Uses Graph RAG (Neo4j) to find suspicious network patterns and builds a case for fraud
• Agent 2 (Defense): Searches for legitimate explanations and challenges the prosecutor’s claims
• Agent 3 (Judge): Reviews both arguments and makes the final verdict

The Live Demo:

1. A suspicious transaction streams in via Kafka
2. Prosecutor Agent queries the Neo4j graph and constructs a fraud case using Gemini
3. Defense Agent counter-argues with alternative explanations
4. Judge Agent renders a verdict with confidence scoring
5. I pop the hood and show: the exact prompts, the Cypher graph queries, and the decision logic

Technical Deep Dive:

• How I structure multi-hop graph context for LLM reasoning
• Prompt engineering to prevent agent “hallucination” on graph data
• Latency battles: why I moved from Gemini Pro to Flash and added prompt caching
• The surprising failure modes: when agents agree too quickly vs. when they hallucinate connections