A research-to-execution loop that gates every step
The problem
Autonomous trading systems fail quietly. A strategy overfits to a thin backtest, the operator sees a good-looking equity curve, and capital goes to work behind a number that was never real. The system had to make that failure mode loud.
What we built
A Python vectorbt research layer proposes and a TypeScript engine disposes. The two share a parity harness so a strategy behaves identically in backtest and live, and the harness fails the build when they diverge.
Every candidate configuration passes through a Deflated Sharpe Ratio gate that accounts for how many configurations were tried. In the first full sweep, 3 of 981 survived. The other 978 were selection noise wearing a good chart.
A per-market strategy arena runs candidates as zero-capital telemetry, promotes survivors to shadow trading on live prices, and only then presents them for arming. Promotion is automatic. Arming never is.
A self-run audit of the execution path produced six critical and four high-severity findings on order sizing, idempotency, and position ownership. All ten were fixed and shipped as reviewed pull requests before that code touched capital again.
Live execution sits behind a kill switch, per-session execution modes, and typed confirmation phrases. The default state of every strategy is off, and nothing in the automation can change that.
What this demonstrates
Autonomous systems that hold real risk, with promotion pipelines, evidence gates, and an off switch the automation cannot reach.
Get in touch
Tell us what you are trying to build and what is in the way. We will tell you early if we are not the right fit.