# Hunch

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Jose Antonio Alatorre](https://cerebralvalley.ai/u/jaa)
- **GitHub:** https://github.com/ms-private-projects/HunchAPI
- **Demo video:** https://youtu.be/8z7mYjCD558
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1521

Hunch turns a thesis into a portfolio you can hold, measure and stand behind.

An investment idea and the portfolio that expresses it usually drift apart: the reasons live in a note, the positions live at the broker, and nobody checks whether one still justifies the other. Hunch keeps them together. The thesis is the unit of record: every position cites its evidence, and every weight, review, proposal and order traces back to the idea.

A person writes a take ("AI data centres will strain the power grid") with their constraints: horizon, cash buffer, single-name cap, maximum positions, exclusions and what would change their mind. The Hunch Agent, on Muse Spark 1.3, critiques the thesis into claims, assumptions and counterarguments. Muse Wonder 1.0 researches it against a universe of US-listed equities and returns claims, beneficiaries and losers, and 3 to 20 candidates, each with a role (pure play, supplier, enabler, hedge, defensive), a case, risks, a conviction and cited evidence, inside a strict JSON schema. Code, not the model, turns candidates into weights under the guardrails and builds a model portfolio whose record starts the day it goes live. Every trading day a Muse Spark review decides hold, rebalance or thesis change, with the triggered conditions and a confidence.

Money moves only through a rebalance proposal the person approves, or that passes an autopilot check inside limits they can see, and only at their own broker (Alpaca or Interactive Brokers) through its connector. Hunch holds no broker credential. Takes can be published, forked with their lineage intact and ranked on a leaderboard by return since activation.

From 0:40 on, the demo video is real screens of the Hunch site on a local stack: the thesis critique arriving from Muse Spark 1.3 (27.7 s), research 1 by Muse Wonder 1.0 (42.3 s: 6 claims, 8 candidates with roles, 65 cited evidence items), the basket the guardrails built from it (two names held at the 15% cap, 10% cash), a real Alpaca paper account linked through the hosted connector, a $10,000 funding whose eight orders filled on Hunch's simulated venue, the daily review Muse Spark 1.3 answered in 17.5 s (hold) and the published take on the leaderboard, with a chapter on how Hunch uses Muse.

What is real today and what is not: research, critique, the daily review and the Agent run live on Muse through the Meta Model API, with every call a recorded request; weights, limits and orders are decided in code. Daily prices on the stack are synthetic until a market-data provider is connected. The hosted Alpaca connector links the paper account and receives the proposal, but orders fill on Hunch's simulated venue until the account's broker keys are stored on that connector, so returns and the leaderboard are not market data yet. The README's "What is real and what is simulated" table says exactly this, and Hunch gives no investment advice.

Hunch is a FastAPI application with an MCP server, one TAU Harness Agent and scheduled Jobs, built and run on the Main Sequence platform. The repository README opens with what was built before and during the hackathon (73 of 92 commits, 29% of all lines, recomputed from git blame by a script in the repo), a judges' guide mapping each criterion to its evidence, and a five-minute check: clone, uv sync, uv run pytest (189 tests: 186 pass with no credentials, 3 skip).

## More from Meta Global AI Developer Hackathon

- [ProtocolEye](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1518)
- [SignBridge AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1519)
- [Shadow Fenced AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1520)
- [Memiti Dapash](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1522)
- [meta_hackathon_winners](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1523)
- [ProbeForge](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1524)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1521. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt