What we built
- Ranking modelA LightGBM ranker trained to distil a professional ratings feed, retrained from scratch every Sunday by a scheduled job with versioned artefacts.
- Win modelA classifier trained on actual results, holdout AUC 0.705, used as a confirmation signal rather than a headline.
- Explainable scorerA deterministic, rule-weighted tips scorer that weighs sectionals, speed map, distance, track, conditions and connections, and writes a reasoning line a reader can argue with.
- Price feeds with fallbackPrimary bookmaker feed, then a best-odds sweep, then a model fair price; exchange prices polled from a small device on a residential connection.
- Engine board and appsAn interactive engine with sort views and a performance API, a native Swift iOS app and a Flutter app, speed maps, and a cloud fallback so the daily picks run even if the main server does not.
- Public siteA static Cloudflare Pages site with a waitlist in KV and a live scoreboard pulled from the performance API, plus a content engine producing tip cards, voiceovers, social posts and a Telegram bot.
On results
We do not publish return-on-investment claims for this work, and we recommend clients do not either. The first version of the performance dashboard overstated results because of look-ahead in the data; finding and fixing that is part of what we deliver. What the public scoreboard shows is every tip, settled, losses included. The engineering lesson applies far beyond racing: a model evaluation is only as good as its leakage audit.
Usage analytics without a tracker
A self-hosted site-stats page reports unique visitors, week-on-week growth, 42-day charts, click-throughs, referrers and devices, with no third-party analytics script and no cookie banner. The same component is available for any product we build.