Falcone AI

Falcone AI is a multi-agent forensic document intelligence system designed for financial crime analysis. Named after Giovanni Falcone — the Italian anti-mafia judge who built cases methodically and never submitted until the evidence was solid. Most RAG systems retrieve and summarize. Falcone interrogates. Upload a case document, ask an investigative question, and five specialized agents go to work: - Il Capo routes the query and compiles the final verdict - Archivist extracts entities, events, and obligations from the document - Chronologist builds a timeline and flags behavioral anomalies - Legalist maps findings to EU legal frameworks via semantic search - Advocatus Diaboli stress-tests every violation claim — invalidating weak findings before they reach the report The Legalist-Advocatus debate loop is the core innovation: Advocatus challenges each finding, Legalist rebuts with fresh retrieval, and the loop runs until the evidence holds or collapses. This produces legally rigorous output rather than hallucinated compliance. The legal corpus covers 7 EU/international frameworks: PIF Directive 2017, AML Directives 2015 and 2018, Market Abuse Directive 2003, Transparency Directive 2004, CONSOB/TUF Italian securities law, and the UN Palermo Convention — embedded with BGE-M3 and stored in Qdrant for high-precision retrieval. Tested on the Parmalat financial fraud case: Advocatus invalidated 6 out of 7 mapped violations on temporal and jurisdictional grounds — correctly identifying that AML Directive 2018 cannot apply to conduct ending in 2003. One validated violation remained, with a referral recommendation to Italian criminal courts and the US SEC. Built with LangGraph for agent orchestration, Qwen2.5-72B-Instruct for reasoning, and deployed on AMD MI300X (192GB HBM3) via vLLM with ROCm.