Band Risk-Review

A multi-agent compliance workflow for high-stakes financial decisions — four specialized agents collaborating through Band.

Submission for the Band of Agents Hackathon (lablab.ai), Track 3 — Regulated & High-Stakes workflows. Built by Ruiyang Zhang (passed all three CFA Program exams). MIT.

The idea

Most "AI finance agents" are a single model answering every question with the same false confidence. Real regulated finance doesn't work that way — it runs on separation of duties: the desk that proposes a trade is not the desk that approves it, and risk numbers come from a deterministic system, not someone's gut.

Band Risk-Review encodes exactly that as a band of four agents, each with one job, collaborating through Band:

1. Data Agent — retrieves the portfolio + market snapshot (MongoDB Atlas via MCP in the real build) and posts a structured evidence packet. Never invents data.
2. Risk Agent — computes deterministic portfolio risk: volatility-weighted risk shares, HHI concentration, a parametric drawdown estimate.
3. Calibration Agent — turns fused signals into a calibrated probability + explicit confidence, sizes the position by fractional Kelly, and abstains when confidence is below floor or signals conflict.
4. Reviewer Agent — an independent compliance gate. It reads the proposal and the risk metrics, checks them against policy limits, and returns one of: APPROVED, APPROVED_WITH_CONDITIONS (trim), ESCALATE (to a human risk officer), or ABSTAIN_UPHELD. It can veto.

The differentiator for Track 3: the Reviewer is a separate agent that can override the Calibration Agent, and every step is an auditable message on the shared Band room — a regulator-friendly trace of who said what and why.

Honest design choice: deterministic core

All risk math lives in risk_core.py (Brier/ECE, fractional Kelly, HHI, drawdown). The agents call it and the LLM only narrates the results. The model never produces a risk number out of its head — the anti-hallucination guarantee that matters in regulated finance.

Run it (no network, no credentials)

python run_demo.py        # four scenarios through the four agents
python test_workflow.py   # 16/16 — math + multi-agent handoff

The demo shows the full range of honest behaviour:

Ask	Outcome	Why
Add to BTC?	APPROVED, size 11.4%	clean signals, drawdown contribution within gate
Add ETH (hot on socials)?	ABSTAIN	funding vs momentum conflict → confidence below floor
Add DOGE?	ABSTAIN	no data on the asset
Add SOL?	ESCALATE	drawdown contribution 9% of NAV > 6% gate → human review

Live on Band (verified ✅)

This is not just a mock — the four agents have been run live as Band Remote Agents, collaborating through a real Band room and powered by Gemini:

python band_agents.py data         # 4 processes (see SETUP_BAND.md for creds)
python band_agents.py quant
python band_agents.py calibration
python band_agents.py reviewer
# then in a Band room with all four added:  @Risk-Data Should I add to my BTC position?
python post_question.py "Should I add to my BTC position?"   # or drive it headlessly
python read_transcript.py reviewer                            # print the room's audit trail

Each agent (band_agents.py) is a custom Band SimpleAdapter: it calls Gemini for intent parsing + narration and the deterministic risk_tools.py for every number (the LLM never invents a risk figure). The handoff Data → Quant → Calibration → Reviewer → user happens entirely through the shared Band room via @mentions — that agent-to-agent collaboration is the judged feature. A verified run of all four scenarios (APPROVED / ABSTAIN×2 / ESCALATE) is in DEMO_TRANSCRIPT.md.

The offline LocalBandBus in band_bus.py remains as the credential-free test double (run_demo.py / test_workflow.py).

Files

Offline scaffold (runs now, no creds — proves the logic):

• risk_core.py — deterministic risk + calibration math (pure functions)
• band_bus.py — the collaboration layer (Bus protocol + LocalBandBus)
• agents.py — the four agents (mock-bus version)
• seed_data.py — sample portfolio + market snapshot
• run_demo.py — orchestrates the four scenarios
• test_workflow.py — tests (math + workflow), 16/16

Live on Band (Gemini-powered — see SETUP_BAND.md):

• risk_tools.py — the deterministic core exposed as agent tools
• band_agents.py — the four agents as Band Remote Agents (Gemini + @mention handoff)
• post_question.py — headlessly post a question into the room (drives the demo)
• read_transcript.py — print the room's audit trail
• agent_config.example.yaml, .env.example — config templates (gitignored when filled)
• SETUP_BAND.md — step-by-step to stand up the four agents
• DEMO_TRANSCRIPT.md — a verified live run of all four scenarios

Built by Ruiyang Zhang (passed all three CFA Program exams). MIT.