All projects
Investing · Multi-Agent

QUORUM — AI Investment Committee

An AI investment committee that argues before it decides — and never makes up a number.

The problem

Almost every "AI invests for you" demo fails the same way: the model states a number with total confidence and the number is wrong — a made-up P/E, a misremembered drawdown, a hallucinated return. In investing, a confidently wrong number is worse than no answer.

The fix is architectural: split the system in two — deterministic computation underneath, language on top — so the failure mode is structurally impossible.

Architecture

  1. 1

    Universe + screen

    51-name, 11-sector investable universe; point-in-time shortlist of the decision-relevant names.

  2. 2

    Research briefs

    Per-name evidence assembled from the tools layer.

  3. 3

    Bull & Bear

    Research independently, argue their case, then rebut each other — genuine disagreement, not one prompt nodding along.

  4. 4

    Macro strategist

    Adds regime context.

  5. 5

    Risk officer

    Computes the downside (vol, beta, VaR, drawdown) and can veto.

  6. 6

    Portfolio manager

    Synthesizes the debate into actual weights.

  7. 7

    Critic

    Stress-tests for groupthink — then the committee loops or converges.

  8. 8

    Memo + human gate

    Allocation, rationale, and surviving dissent; streamed live to the Committee Room over SSE; a paper portfolio advances daily.

Key tradeoffs

Determinism boundary — prices (yfinance), fundamentals (SEC EDGAR), macro (FRED), and risk (NumPy/SciPy) are computed in Python; a grounding guardrail rejects any unsourced number; the LLM only narrates.

Why · This is the whole point: it kills the confident-wrong-number failure that ruins AI-investing demos.

Bull and Bear research independently before they rebut.

Why · Genuine disagreement has to be engineered — a committee that always agrees is just one opinion in six hats.

Honest backtest — point-in-time, no lookahead, trading costs included, vs SPY; reported as directional.

Why · Backtests are small-sample and regime-dependent; better to under-claim than oversell an alpha.

Runs with zero keys on free data, with provider failover on top.

Why · Robustness and $0 — a single rate limit never takes the committee down.

Eval results

vs SPY
Backtest

$10k paper book, point-in-time (no lookahead), trading costs included — reported as directionally reasonable, never "beats the market".

deterministic
Numbers

Every figure computed in Python; a grounding guardrail rejects unsourced numbers before they enter the debate.

live daily
Track record

The paper portfolio is advanced by a scheduled GitHub Action, accumulating a real history.

Production proof

The artifact that keeps the numbers honest — the eval harness / monitoring gates that run in CI, not a one-off notebook result.

Honest-by-construction backtest

CI · passing
Lookahead point-in-time
Trading costs included
Every figure sourced chip

Results are reported as directionally reasonable, never as an alpha claim; low-confidence decisions are flagged, not hidden.

Demonstrates the architecture finance AI actually needs — verifiable math underneath, language on top — applied to the hardest possible audience: the markets.

Let's talk

I'm focused on finance AI — credit risk, RegTech, AML, and agentic investment research. Open to roles, mentorship, and collaborators in fintech, quant, and bank AI.