01 · The work
Refract
A quantitative research and trading platform. It turns the way a skilled chart reader works into features a machine can measure, tests strategies against twenty years of market data, and runs a paper book on Interactive Brokers every trading day.
What it is
Refract started as a question: can the way a good discretionary trader reads a chart be written down as features, and then be tested honestly? It grew into a platform. A data layer collects prices, options and macro series and checks its own coverage. A feature pipeline only ever sees what was known at the time. A backtester prices fills from the real tape. A live paper book trades the result.
How it runs itself
The trading server runs the market day on timers. Every job leaves a verdict, pass, fail or skip, and silence counts as a failure. When something breaks, the alert wakes an agent on the server that reads the logs and the data store, ships the fix and reports back. A daily sweep looks for the failures nobody was told about, and for the alerts that cried wolf.
How it decides
Strategy changes go through an experiment queue. Variants are proposed, scored by a parallel backtest engine, and adopted only when they clear a bar set before the run. The live book is then held to the backtest's promises: fill rate, credit captured, budget deployed, win rate and drift are measured every session against the benchmark's own distribution.
What it taught us
The hard part was never the model. It was keeping the numbers honest: drawdowns measured on capital rather than on a swollen equity line, fills priced from what actually traded, and a commit hook that refuses any message claiming a green test run that never happened. It trades a paper account. Real money waits on a go-live ledger that is not yet green.