The AI Bookie

The problem

Sportsbook prices hide a built-in margin and the books disagree with each other. Bettors have no transparent, evidence-graded view of the consensus. The site explicitly does not claim to predict outcomes, and never places a bet.

What it does

• Removes each book's margin and turns the consensus into a 0–100 score with an A+ to F grade.
• Rejects incomplete markets, and measures best-line value only against the other books.
• A daily board that locks each morning, an evidence explorer and a pick archive.
• A "Brain" view that shows scan integrity, data source, capture time and the next refresh.
• Email digests by sport, market and day, with double opt-in.

How it's built

• Vinext (a Vite-based, Next.js-compatible framework) with React, deployed as a Worker with an hourly cron.
• D1 through Drizzle; odds and sports data from The Odds API, API-Sports and the MLB Stats API.
• Scoring is deterministic market math, not a model's guess.

Engineering decisions

• A hard data budget: each scan is frozen for 24 hours, a reserve of provider credits is never touched, and daily spend is capped and reconciled with what the provider bills.
• When data or budget isn't available, it serves the last verified board with an explicit warning and a reason code. Only a complete, current scan can create the official record.
• There is no code path that places a bet, every feature flag defaults to off, and a verifier scans the code for betting actions and credentials.
• Release gates are tested against deliberately broken copies of the code, which caught a real bug in the verifier itself.
• Append-only ledgers, guards against using future data, and checks that a re-run gives identical results.