Work · Autonomous system
Alphie
An AI-traded sleeve inside deterministic rails, on NSE stocks.
What it is
Claude trades a discretionary sleeve of NIFTY 50 stocks from a small GCP VM. A single session runs each trading day, resumed once a minute from 09:15 to 15:15 IST, carrying the whole day in memory; it reads the tape, its own rolling record and the day's rails, then manages at most one bracket trade: a stop 0.7% from entry, and once the trade is 0.2% in profit a trailing stop arms and tightens toward the price every five seconds. Server-side code bounds every number it picks: notional per trade, one position at a time, trades per day, a daily loss halt and a cumulative kill switch that only I can reopen. The backtested gap-fade strategy Alphie started as still runs its whole pipeline each morning, on paper only, so its evidence keeps accruing without money on it.
What was hard, or unusual
- The model picks the trade, tested code picks the limits. Rails run server-side and fail closed, so a session that wants more cannot argue its way past them.
- Every session closes itself: a data snapshot committed to the repo, an emailed brief, and a critique written by a scheduled Claude Code routine. The evidence loop runs whether or not I read it.
- Gap fade was demoted to paper rather than deleted, so its shadow experiments keep scoring against simulated fills while the sleeve carries the risk.