WhistleBets

Everyone argues about the refs and almost nobody keeps score. I wanted to see how fast one person could stand up a real data product with AI as a teammate, and whether the claims would survive a proper test. Monty, the zebra in the referee shirt, keeps it from taking itself too seriously.

• Tonight's crew and what their history shows, published before the game
• The sample size sits next to every number, so you can see how much to trust it
• Career tendency pages for officials, a nightly slate and a weekly report
• NBA live now, NHL being added on the same core

• One model writes, another judges. Claude Haiku drafts each read, then Jev, a typed-decision model from TypeSafe AI, screens it against eight hard rules
• Fully automated publishing behind fixed wording checks, a kill switch and a weekly digest. In the first week 132 lines went out and one was held
• Jev can't change a read until its calibration is published, so a gate has to show its record before it gets a say
• One shared core with an adapter per league, so a new sport is a new adapter and not a new product

• Pre-registered test on an unseen 2025-26 season: the foul tendencies held up, with a t-statistic of about 3.3
• Same test, the honest half: no game-totals strategy beat the bookmakers' margin. Every game in the crew's direction over 8,331 bets won 51.2%, a loss of 2.1%
• From first commit to a working site in three days

Three days to a working site.

3 days · 285 commits · 14,183 games

The hard part was staying honest. Data like this wants to become tout talk, so I wrote the test down before I ran it and let it come back with the answer it came back with. Referees really do have tendencies, and that alone doesn't beat the market. The site says what the numbers can carry and stops there.

1. Models

   • Claude Haikuwriting
   • Jev by TypeSafe AItyped decisions

2. Build

   • Next.js
   • React
   • TypeScript
   • Zod
   • MCP server

3. Data and hosting

   • Vercel AI Gateway