ANT-ZIP

An evidence-first microcap research colony that exposes entry timing, risk, and position-aware exit zones.

Comment

Inspiration

Microcap traders face a flood of noisy price moves, thin evidence, and emotionally charged claims. ANT-ZIP turns that noise into a transparent research workflow: every candidate must pass visible, role-specific gates before it appears.

What it does

ANT-ZIP is a local-first U.S. microcap research colony with four specialist agents:

  • Queen Ant accepts only the highest-conviction candidates that pass every quality, freshness, catalyst, liquidity, and risk gate.
  • Soldier Ant tracks constructive early formations and shows which strict gates are still missing.
  • Beast Ant evaluates controlled pullbacks and possible continuation structures after a prior expansion, while vetoing exhausted setups.
  • Doctor Ant combines a ticker with the user's entry price to show current P/L, position timing, an invalidation reference, and staged exit reference zones.

Each visible candidate includes an entry-timing state, entry range, stop reference, staged exit references, evidence, and risk warnings. If no fetched stock passes a role's complete gate, that chamber stays empty instead of inventing a recommendation.

How we built it

The submission is a portable Python 3.11+ application using the standard-library HTTP server, a deterministic domain model, vanilla JavaScript, and responsive CSS. Yahoo Finance daily history is accessed through yfinance for live research, while a deterministic four-ticker fixture provides an explicit fallback for reliable judging. No OpenAI API key, broker connection, or server-side JavaScript execution is required.

The same normalized market snapshot drives classification and Doctor research zones, keeping the interface and evidence trail consistent. The repository includes automated tests, a privacy scan, portable setup instructions, and English submission documentation.

How we used Codex and GPT-5.6

Codex with GPT-5.6 was the primary development partner throughout Build Week. It helped challenge and narrow the product scope, design the four-role gate model, rebuild the portable Python architecture, implement the local server and responsive interface, generate and process character sprite sheets, write regression tests, diagnose browser behavior, and verify the final submission package.

I retained the key product decisions: the four ant roles, strict empty-state behavior, local-first runtime, position-aware Doctor workflow, and the requirement that simulated, live, and unavailable evidence are never presented as the same thing. The primary Codex build task is documented with Session ID 019f5e3a-fe63-7181-99a2-29301c67d80c.

Challenges

The hardest problem was balancing an engaging game-like interface with honest financial evidence. We deliberately avoided guaranteed language and designed explicit vetoes, data-age labels, limitations, and empty states. Another challenge was creating a portable submission that runs without secrets while still demonstrating live behavior.

Accomplishments

  • A working local-first product rather than a static mockup
  • Four understandable specialist roles with visible decision boundaries
  • Position-aware ticker analysis using current price structure and the user's entry price
  • Deterministic fallback data for reproducible judging
  • 46 automated tests plus privacy and JavaScript syntax checks
  • A responsive, animated anthill interface using only local runtime assets

What we learned

Agent collaboration works best when product decisions remain explicit and verification is treated as part of the feature. Codex accelerated implementation most when the acceptance criteria were observable: a candidate either passes a gate, a zone ordering invariant holds, or a browser interaction succeeds.

What's next

Future work could add licensed filings and news sources, saved watchlists, configurable risk profiles, and broker-independent alerts. These would remain optional evidence sources; the core local workflow would continue to work without paid APIs.

Safety boundary

ANT-ZIP provides research references, not predictions, guarantees, or individualized financial advice. Yahoo Finance data may be delayed or unavailable, and the application does not place trades.

Built With

Share this project:

Updates