Proud to share QuantPilot, my solo build for the Open Agent Hackathon 2026 — an autonomous pre-market options research agent, built in the 6-day build window.
The system: every morning before the open, it scores 12 names on a liquidity board (open interest, options volume, momentum, relative volume), detects order-block zones and liquidity pools, reads market breadth and sector flow, then runs a structured bull vs bear debate with a judge verdict. Every number on a trade card traces back to a data row, enforced by a grounding verifier.
Under the hood:
- Data: Zetaris AI Data Harness as a single SQL interface over 46 REST tables for quotes, bars and indicators; options chains direct from Yahoo
- Reasoning: NVIDIA Nemotron (ultra-550b) through Meterless Nightshift as the primary provider, with per-call model logging
- Guardrails: delta floor, spread caps, IV-rank and earnings/event gates, 1% account risk sizing, and a deterministic risk-review pass that can only downgrade a call
- Memory and audit: vendored H-Mem engine recalling past calls, an append-only SQLite journal, and a 2-loss circuit breaker
- Verification: 268 tests passing; live runs Oct 23 and Oct 26 with zero errors; grounding 113/115 citations verified on the Oct 23 run and 91/91 on Oct 26; an impact study over 6,504 live contracts
The result was honest, not a win: 66.5/100, rank 252 of 547, and no advance to the final round. Innovation scored 6.5/15 — the winners tackled broader, verifiable problems. But the demo scored 12.5/15, above all three winners in my track, and the engineering held up under live market data.
Repo (public): https://lnkd.in/emZA367k
Live explorer: https://lnkd.in/eZJZx9jp
Certificate of participation attached. Next time, the bigger problem comes first.
#AIagents #Hackathon #OpenSource #LLM #FinTech
Top comments (0)