Quantis AI Trading
An autonomous AI agent trades a US-stock paper account with volatility-targeted leverage, a small mechanically-scored single-name sleeve, and a standing protective-put hedge — every position wrapped in hard survival mechanics the LLM cannot override.
Personal engineering project · Paper trading only · Not investment advice
A volatility-targeted trader that scales leverage to how turbulent the market is — more when it's calm, automatically less when it isn't. An LLM runs the routines, but the core's leverage is a formula (target ÷ realised volatility, capped 2×), the single-name sleeve is mechanically scored, and a standing protective-put hedge rides underneath. The vol-target is the engine; the drawdown ladder, VIX kill-switch, margin-distance gate and put hedge are the defence. Survival is rule one — a wipeout can't compound.
How it works
Size to volatility
An Opus pre-market routine measures QQQ's realised volatility and sets the core's leverage by formula: target ÷ realised vol, capped at 2.0×. Calm tape earns more leverage; turbulent tape automatically gets less — before VIX even spikes. Below the 200-day moving average the exposure is halved. The number is a calculation, not a vibe — and it backtests to a higher Sharpe than holding QQQ outright.
Deploy the core + the hedge
A leveraged QQQ/SPY core is sized to the vol-target and gets a catastrophic backstop stop. A standing protective-put hedge on QQQ is kept on (≈10–15% OTM) so a gap can't spiral the book. A small single-name sleeve (≤15% of equity, unlevered) is sized by a mechanical score — momentum, earnings revisions, sector strength — only for names already trending up; the LLM can veto, never upsize.
Cut fast, let winners run, survive
Two intraday risk-checks (10:30 + 13:30 CT) plus midday hunt for trouble. Single-name losers are cut fast at −1.5×ATR or −7%; winners move to breakeven, bank a third at +1.5×ATR, then ride a Chandelier trail. Survival runs continuously — drawdown ladder (−10/−20/−25%), VIX kill-switch, margin-distance gate, fast-reversal trigger, and a de-lever to flat into every close. A deadman watchdog flattens the book if the routines ever stall. Friday's Opus review runs an alpha/beta attribution and the kill-check.
How it's built
No databases, no ORM, no in-memory state. Every memory file — strategy rulebook, trade ledger, daily research, the atomic risk-state file, weekly reviews — is committed to a private GitHub repo. Every routine run is a fresh container that clones, decides, commits, exits.
Hard rules the LLM cannot override
Leverage is set by formula and held inside a cage of survival mechanics enforced before any order is placed. The LLM runs the routines and picks the single names; the rules derive the core's leverage from realised volatility, cap it at 2×, keep the hedge on, and pull risk the moment drawdown or volatility crosses a line.
- Volatility-targeted leverage — gross = target ÷ realised vol, hard cap 2.0×
- Trend filter — leverage halved when QQQ is below its 200-day MA
- Single-name sleeve capped at 15% of equity, unlevered, sized by a mechanical score
- Trend-confirmed entries only — above the 20-day MA with positive momentum
- Cut losers fast — exit at −1.5×ATR or −7%, whichever hits first
- Let winners run — breakeven move, scale out a third at +1.5×ATR, then a Chandelier trail with no fixed cap
- Standing protective-put hedge on the core (premium = max loss, ≤3% of NAV per year)
- Catastrophic stop on the leveraged core; volatility-adaptive ATR stops on every single name
- Drawdown ladder: −10% halve · −20% go flat · −25% hard halt; VIX kill-switch + 15% margin-distance gate
- De-lever to flat into every close + weekends; margin financing modelled; deadman watchdog flattens on a stall
What this actually is
This is a learning project. The infrastructure — an autonomous LLM agent, hard-rule risk discipline, git-as-memory, an atomic risk-state file — has engineering value regardless of P&L. There is no signal service, no subscription, nothing for sale, no community — just a private dashboard for the operator and a public page explaining what was built.
See it run
The dashboard is private. The only person with access is the operator. Click below if that's you.
Sign in