A panel of AI agents puts every disputed insurance claim on trial — and dynamically recruits a sixth investigator the moment fraud is alleged. The debate becomes the legally-defensible audit trail. A human officer keeps the final word.
Band of Agents Hackathon · lablab.ai · Track 3: Regulated & High-Stakes
- 🔗 Live app — recourseband.duckdns.org
- 🎬 Commercial — youtu.be/io2CgqYpSek
- 🖥️ Live walkthrough — youtu.be/kY6AXSlKqhk
- 📑 Pitch deck —
deck/deck_v4.pdf
The live adjudication room: five agents debate, the Coordinator recruits Quinn (SIU) when fraud is alleged, and a signed verdict lands — a real run on the production app.
Crestview Mutual (a fictional insurer) denied a claim. Instead of one overworked reviewer deciding the appeal alone, Recourse convenes a panel of specialist agents in a single Band room — five standing, plus a sixth investigator on call — and lets them argue it out:
| Agent | Role | Job |
|---|---|---|
| Coordinator | Orchestrator | Opens the case, convenes the room, routes every handoff |
| Blake | Claims Evaluator | Argues the merits — builds the case for the insured |
| Morgan | Policy Analyst | Cites the exact governing clauses (§7.3, §12.1 …) via RAG |
| Alex | Devil's Advocate | Fights to deny — so no weakness goes unexamined |
| Sam | Resolution Notary | Writes the final, signed, reasoned resolution |
| Quinn ◉ | Special Investigations Unit | Not a standing panelist — the Coordinator recruits Quinn into the live room (Band add_participant) only when fraud or misrepresentation is alleged, so a claim is never denied on unproven suspicion |
Dynamic agent discovery — Quinn is summoned mid-debate, on demand, not pre-wired into every case. When no fraud is alleged the panel stays at five; when the denial rests on a fraud/misrepresentation claim, the Coordinator pulls in the investigator to test it.
The entire conversation is persisted, ordered, and SHA-256 hashed — tamper-evident on the record. A human claims officer then approves or overrides the resolution, and the closed case is downloadable as an officer-signed, SHA-256-stamped Adjudication Record — printable to PDF, plus a machine-verifiable JSON export.
 Dynamic recruitment — the Coordinator pulls Quinn (SIU) into the live room, mid-debate, only when fraud is alleged |
 Signed verdict — APPROVED with cited clauses, the SIU finding, and the human officer's sign-off |
 Signed record — the officer-signed, SHA-256-stamped record, printable to PDF |
 JSON export — the full transcript + SHA-256 hash, machine-verifiable |
Each case also carries clickable evidence (police / fire-marshal reports, scene photos) the panel reasons over — so nothing is invented.
Recruitment is discriminating: a clean dispute runs five agents; a fraud allegation pulls in the sixth. Two cases are seeded pending so anyone can run a full debate end-to-end.
| Case | Claim | What happens |
|---|---|---|
| David Chen | $12,000 collision, denied | pending · run it live — 5 agents, no fraud → Quinn not recruited |
| Lisa Park | $4,200 theft, denied as undisclosed commercial use | pending · run it live — 6 agents, fraud alleged → Quinn recruited live |
| Marcus Reyes | $9,300 fire, denied as a staged loss | closed showcase — pre-adjudicated, full 6-agent transcript, officer-signed |
- Frontend — Next.js 14, TypeScript, Tailwind; neo-brutalist "Verdict" design.
- Backend — FastAPI, async SQLAlchemy, SSE; a Coordinator-driven turn engine.
- RAG — policy clauses embedded with
all-MiniLM-L6-v2(384-dim), pgvector cosine search. - Agents — a long-lived worker (
agents.run_agents) keeps the 5 standing Band agents connected, plus Quinn on call (recruited into the room only when fraud is alleged).
Postgres runs on host port 5433 locally (to avoid clashing with a native install).
cd backend
python -m venv .venv && .venv\Scripts\activate # Windows
pip install -r requirements.txt
docker compose up -d db # from repo root: Postgres 16 + pgvector (5433)
.venv\Scripts\python database/seed_data.py # seeds the 3-case demo (2 pending + 1 closed w/ Quinn)
uvicorn main:app --reload # http://localhost:8000/api/healthcd backend && .venv\Scripts\python -m agents.run_agentscd frontend
npm install
npm run dev # http://localhost:3000Copy backend/.env.example → backend/.env and fill in the Band / AI-ML / Featherless keys.
cd backend
pip install -r requirements-dev.txt
python -m pytest tests/ -q # 40 unit tests — fraud-recruitment gate, payout, SHA-256 hash, verdict parsingThe suite covers the logic that must be provably correct: _alleges_fraud (the SIU-recruitment
trigger), the deterministic payout, the tamper-evident transcript hash, and verdict parsing — all
pure functions, so it needs no database, Band connection, or network.
Production runs as a Docker Compose stack on a VPS (db · backend · agents ·
frontend · caddy), behind OpenLiteSpeed as the public reverse proxy with
Let's Encrypt TLS and a DuckDNS domain. The stack lives in deploy/:
cd deploy
cp .env.example .env # fill in secrets
docker compose up -d --build
docker compose run --rm backend python database/seed_data.py # seed the 3-case demoQuinn is feature-flagged. Set
BAND_QUINN_AGENT_ID/BAND_QUINN_API_KEY/BAND_QUINN_HANDLEindeploy/.envto enable live recruitment of the 6th agent. Leave them blank and the panel runs as the standing five — zero risk, no code change.
DEPLOY.mddocuments an alternative managed-PaaS path (Supabase / Railway / Vercel); the live deployment above uses the self-hosted Docker stack indeploy/.
- Scaffolding, database (schema, models, RAG over pgvector, seed)
- Band SDK + Room Manager + Coordinator orchestration
- The 5 standing agents (Coordinator + Blake, Morgan, Alex, Sam)
- Quinn (6th agent) — dynamically recruited mid-debate when fraud is alleged
- FastAPI backend + debate orchestrator + SSE
- Frontend (dashboard, live debate room, resolution panel, signed-record download)
- Unit tests — 40 passing: fraud-recruitment gate, payout math, SHA-256 audit hash, verdict parsing
- Deployed live + demo video + pitch deck
MIT