← /projects

RaceCast

Four AI models — Claude, GPT, Gemini, and Grok — compete to predict every Formula 1 race of the season. Scores update against real results on a season-long leaderboard.

Stack

  • TanStack Start
  • React
  • TypeScript
  • oRPC
  • Drizzle ORM
  • Postgres
  • Vercel AI SDK
  • OpenRouter
  • Tailwind
  • Radix UI

Overview

Four LLMs — Claude, GPT, Gemini, and Grok — compete to predict every Formula 1 race of the season. Pre-race picks are scored against real results and roll into a season-long leaderboard, settling the "which model is smartest" debate race by race.

  • Pre-race predictions from 4 models with distinct prompts
  • Side-by-side comparison of each model's picks
  • Accuracy scoring — winner, podium, top 10, fastest lap
  • Live race dashboard — weather, tires, pit stops from OpenF1

What it solves

Every time a new LLM drops, the internet argues over which one is actually the smartest. Benchmarks are abstract and hard to feel. Real-world prediction is the opposite — either the model called the podium right or it didn't.

RaceCast turns four models loose on every Formula 1 race of the season. Each one commits picks before the lights go out, the results score them automatically, and the leaderboard grows in public. No curated test set, no cherry-picking — just 24 races of being wrong or right.

Highlights

  • Four models with distinct prompting strategies predicting side by side
  • Consensus indicator flags when all four agree (and when they split)
  • Live race dashboard pulling weather, tires, and pit stops from OpenF1
  • Season + championship leaderboards that update themselves