Yield-paca: Multi-Agent AI Trading System

Yield-paca is an explainable, safety-first AI trading system. Each trade is a structured conversation: Market Data, News, Macro, and Insider agents research in parallel; Bull and Bear debate; a Research Manager creates a typed plan; a Trader LLM proposes an option structure; deterministic RiskManager and PortfolioManager decide whether it reaches Alpaca. MCP is the safety boundary. Only PortfolioManager receives Alpaca account, position, and trading tools. Analysts get read-only tools; the Trader LLM cannot submit orders, even through prompt injection. Swap the MCP broker and the pipeline can run elsewhere. Three strategies share the same gates: - Mid-Band Movers, 11:30 AM ET: 2h, 4h, and 23.5h buckets; ATM calls. - Overnight Reversal, 9:35 AM ET: winners to calls, losers to defined-risk bear-call spreads; closes 15:55 ET. - Bull/Bear Pipeline, every 15 minutes: briefs, debate, synthesis, LLM design, deterministic approval. Temporal runs parent/child workflows across five task queues; schedules survive worker restarts. Re-triggers are idempotent while positions remain open. Postgres retains agent runs, plans, risk decisions, orders, and fills; AlpacaMirror syncs continuously. RiskManager enforces position, Greeks, and decision-TTL limits. PortfolioManager rechecks approval per leg, preventing naked sell-to-open orders on the level-3 paper account. During the Aug–Sep 2026 hackathon run on paper account PA3PR2RMFE9Y, equity ranged from $89K–$98K intraday. The system placed 60+ orders, recorded 22+ fills, and deployed about $13K through Overnight Reversal. Built with Ruby, Rails, Vue, Temporal, Postgres, Claude via ruby_llm, and MCP servers for Alpaca, TradingView, OptionsFlow, and FRED. Next: per-strategy caps, backtesting, options-flow signals, and adaptive DTE selection. Open source: irox/images/trader.