Most retail "AI trading" products ask you to pick indicators. NorthStar asks one question: where do you want to end up? You state a destination in one sentence. The system translates it into required return, computes honest odds from walk-forward backtest distributions and Monte Carlo, and if the goal is unrealistic it says so in red, with alternatives that would actually work - before a single simulated dollar moves. Then it sails: an agentic loop (Google ADK Workflow + Gemini) runs every 15 minutes against an Alpaca paper account. Options-first: a wheel state machine, three defined-risk spreads submitted as one atomic multi-leg order, and a mechanical exit manager. A deterministic 22-rule risk gate sits between every proposal and the broker; LLMs triage, debate and explain, but never touch the money path. Every decision - including every rejection - lands in an append-only journal with reason codes, and the whole account is publicly auditable on a live read-only cockpit. Our contest week closed down 7.04%. One momentum name gave back 85% of that; the gate passed it legitimately, the strategy was wrong, and the journal says exactly that. We traced the root cause and shipped the fix the same week. Honest ledger is the product. Stack: Alpaca Trading API + option chain, alpaca-mcp-server (read-only), Google ADK + Gemini, TimesFM, FastAPI, Next.js, Cloud Run + Firestore.
Category tags: