Riskgate

Riskgate scans a watchlist of liquid underlyings, builds real Bull Put Spread candidates from Alpaca’s live option chain (real strikes, greeks, and quotes), scores them deterministically, and asks an LLM for a structured second opinion. Every candidate passes through a deterministic Risk Engine checking liquidity, DTE, credit, defined risk, daily loss limits, portfolio risk, duplicate exposure, and position sizing before anything can be submitted as a real (paper) order. The AI never has direct access to `submit_order`. The core rule is simple: the LLM proposes, it never decides. Its output is validated against a strict schema before it can influence anything, and even a valid “APPROVE” can be overridden by the Risk Engine. This was verified through live paper-account testing, which exposed and fixed two real bugs: AI calls on hopeless candidates that should have been pre-screened, and a duplicate-exposure check that failed to update mid-scan. Built on `alpaca-py` (Trading + Market Data APIs, including multi-leg MLEG orders), the official Alpaca MCP Server (read-only query/demo path), and Featherless.ai (open-weight models such as Qwen3). The autonomous loop runs continuously across the multi-day scoring window via GitHub Actions, triggered by cron-job.org after GitHub’s own scheduler proved unreliable at short intervals. Every decision is journaled in a shared Supabase Postgres database with full market/option inputs, AI rationale, and itemized risk checks. The same live journal is accessible from GitHub Actions, local development, and a public Vercel dashboard with four pages—Overview, Daily KPIs, Positions & Trades, and Decision Journal. All pages support real date-range, symbol, and decision filters, making the reasoning behind every trade fully reconstructable.