Parallax is an autonomous AI trading agent built for the Alpaca AI Trading Agents Hackathon, implementing a "dispersion barbell" options strategy across two sleeves. Sleeve A harvests rich single-name implied volatility through defined-risk put credit spreads and iron condors on liquid, optionable names, entered only when a name's IV30/HV20 ratio and vega-weighted dispersion score clear a threshold and no binary catalyst is pending. Sleeve B is a small, cheap SPY put debit spread that pays off on large index moves around scheduled macro events, sized as a percentage of NAV. The pipeline is fully autonomous end-to-end: a deterministic liquidity/IV screen narrows a 25-name universe, a Gemini-powered catalyst gate vetoes names with unpriced binary risk, a second Gemini call proposes the specific spread structure, a 13-gate risk engine that can only shrink or reject a proposed trade (never enlarge it) validates every proposal, and execution submits real multi-leg limit orders through Alpaca's Trading API — with every stage, pass or reject, written to an append-only audit journal. A dated endgame state machine enforces the trading calendar and guarantees the account is fully flat before the submission deadline, independent of what any LLM call returns that day. Safety is treated as a first-class design constraint, not an afterthought: paper trading is a hard invariant enforced at every layer, every short option leg must be covered within the same multi-leg order (no naked shorts, ever), orders are limit-only by construction, and a multi-underlying basket reserves its full max-loss capital in the journal before the first leg is submitted. The system ships with 180+ automated tests, including a property test asserting the risk engine can never increase a proposed trade's size across 200+ randomized inputs.
Category tags: