AEGIS: Autonomous Financial-Crime Investigation

AEGIS (Agent-based Evidence-Grounded Investigation System) is an autonomous AML investigation mesh that replaces single-pass LLM guessing with a coordinated team of 15 specialist agents. When a transaction alert arrives, an Intake & Orchestrator agent opens a Band case room and recruits specialists by alert type. Transaction Pattern, Identity/KYC, Network Graph, and External Intelligence agents post structured evidence to the shared room. A Challenger (Red-Team) agent then argues the innocent explanation — any suspicion that survives a real attempt to break it is far more likely real. A Verifier rejects any claim not backed by a cited evidence reference. The Adjudicator synthesises the deliberation into a verdict, confidence score, and evidence chain, then applies risk-based autonomy: benign cases are auto-cleared and logged; genuinely uncertain cases escalate to a human Compliance Officer. Beyond the single institution, a Consortium Liaison shares abstract pattern descriptors with peer-bank agents via Band — zero raw records ever cross institutional boundaries. A Quality Auditor, Strategic Intelligence agent, Org Policy agent, Pattern Memory agent, and Chat Analyst complete the 15-agent roster, providing post-case auditing, cross-case typology detection, per-bank policy enforcement, institutional memory, and natural-language Q&A grounded in casebook data. Accuracy is measured on the public PaySim / IBM AML dataset — external labels the team did not author — on a fixed held-out slice, giving an honest FP-reduction number versus a single-LLM baseline with identical data access. Built on Band SDK · CrewAI · LangGraph · FastAPI · Qdrant · NetworkX · Featherless AI · AI/ML API · React/Next.js. Live at aegis-g7vl.onrender.com.