ConclaveAI

Multi-agent AI system that analyzes Suspicious Activity Reports (SARs) to reduce false-positive escalations, by simulating a compliance review committee through investigator, compliance, behavioral, and adversarial reasoning agents.

Overview

Financial institutions generate large volumes of SARs, and manual review is slow and inconsistent, leading to over-escalation of low-risk activity. ConclaveAI addresses this by running a SAR narrative through a "committee" of specialized LLM agents that debate the case from different angles before a final arbiter issues an explainable risk verdict — mirroring how a real compliance review board reasons through a case.

How it works

A SAR (PDF or raw text) is ingested and passed through a sequential agent pipeline orchestrated with LangGraph:

1. Financial Investigator — analyzes transaction patterns, entities, and flow of funds
2. Compliance Officer — checks findings against regulatory guidance (e.g. FATF, FinCEN)
3. Behavioral Analyst — flags anomalous account/customer behavior
4. Devil's Advocate — builds counter-arguments and stress-tests the emerging consensus
5. Risk Arbiter — synthesizes all findings into a final risk score, suspicion level, and recommendation

Each agent's output streams to the frontend in real time over Server-Sent Events, so the review can be watched turn-by-turn instead of waiting on a single black-box response.

Architecture

• Agent orchestration — LangGraph coordinates the five-agent reasoning pipeline and shared state
• LLM inference — local models served via Ollama, keeping SAR data on-premises rather than sent to a third-party API
• Retrieval — ChromaDB vector store for grounding agents in relevant regulatory/context knowledge during analysis
• Backend — FastAPI service exposing PDF ingestion, entity extraction, streaming analysis, case management, audit logging, and governance endpoints
• Frontend — Next.js + TypeScript dashboard (Live Monitor, Case Files, Audit Ledger, Oversight) for reviewing agent reasoning and outcomes

conclave_backend/
├── api/main.py                  # FastAPI app: SAR upload, streaming analysis, cases, audit, governance
└── services/
    ├── pdf_ingestion.py         # PDF → clean text (PyMuPDF)
    ├── entity_extractor.py      # Persons, orgs, accounts, jurisdictions, relationships (regex + spaCy)
    └── streaming_runner.py      # Drives the 5-agent LangGraph pipeline, emits SSE events

ConclaveAI-Frontend-main/
└── src/pages/dashboard/
    ├── live-monitor/            # Real-time view of the agent committee working a case
    ├── case-files/              # Saved case history
    ├── audit-ledger/            # Immutable, hash-chained log of every agent/analyst action
    └── oversight/                # Governance metrics: bias audit, factual consistency, regulatory adherence

Key features

• Explainable risk scoring — every verdict is backed by the individual findings of each agent, not a single opaque score
• Entity & relationship graph — auto-extracts persons, organizations, accounts, jurisdictions, and amounts from the SAR narrative and maps them into a network graph, with high-risk jurisdictions flagged automatically
• Adversarial review — a dedicated agent argues against escalation, surfacing weaker cases before they reach a human reviewer
• Live streaming analysis — agent-by-agent progress via SSE instead of a single blocking call
• Audit trail — every document upload, agent completion, and analyst decision (confirm/override/escalate) is logged with a verification hash
• Governance dashboard — bias audit (override rate), factual consistency, and regulatory adherence tracked from real case history

Tech stack

Layer	Technology
Agent orchestration	LangGraph
LLM inference	Ollama (local models)
Vector retrieval	ChromaDB
Backend API	FastAPI (Python)
PDF parsing	PyMuPDF
NER / entity extraction	spaCy (regex fallback)
Frontend	Next.js, React, TypeScript, Tailwind CSS

Getting started

Prerequisites

• Python 3.10+
• Node.js 18+
• Ollama installed with a local model pulled (e.g. ollama pull llama3)

Backend

cd conclave_backend
pip install -r requirements.txt
uvicorn api.main:app --reload --port 8000

Frontend

cd ConclaveAI-Frontend-main
npm install
npm run dev

Visit http://localhost:3000 for the dashboard; the API runs on http://localhost:8000.

API endpoints

Method	Endpoint	Description
GET	/api/health	Ollama + backend status
POST	/api/upload-pdf	Extract text from an uploaded SAR PDF
POST	/api/extract-entities	Extract entity graph data from SAR text
POST	/api/analyze	Stream the 5-agent analysis over SSE
GET / POST	/api/cases	List / save analyzed cases
PATCH	/api/cases/{id}/status	Confirm, override, or escalate a case
GET	/api/audit-log	Retrieve the audit trail
GET	/api/governance	Bias, factual consistency, and regulatory metrics