Pitbrain — AI Formula 1 live race intelligence, built on OpenF1

[GCloke]

Hi all

I wanted to share Pitbrain.ai — a Formula 1 site I've been building that leans heavily on OpenF1 for its live and historical data. First off, a big thank you to @br-g and everyone here: OpenF1 is the backbone of the whole live side, and it's been a genuine pleasure to build on.

What it is
Pitbrain takes real session data and turns it into three things: a "what's happening right now" live view, an AI race oracle that publicly forecasts each Grand Prix, and a set of prediction games — all designed to keep people engaged across a race weekend.

How it uses OpenF1
The live experience is driven by the OpenF1 MQTT feed (mqtt.openf1.org). A worker subscribes to:
v1/car_data, v1/location, v1/laps, v1/intervals, v1/stints, v1/pit, and v1/race_control

..and fans those out to browsers over a WebSocket, so the Paddock Live page updates in near-real-time during sessions — positions, gaps, tyre stints, pit stops, race-control messages and weather.

For results, classifications and historical sessions I use the OpenF1 REST API (/v1/sessions, /v1/stints, /v1/laps, /v1/pit, /v1/race_control, etc.), which powers a full session archive you can replay after the fact, plus a race-pace heat-map built from lap times.

AI Race Oracle
About 30 minutes before lights-out, an AI model posts its podium prediction for the race — reasoning over qualifying results, each driver's recent and historical form, weather, and current F1 news. The interesting part: after the chequered flag, every call is scored against the actual result, so there's a public, running accuracy tracker across the season. It's an AI that has to show its work and own its misses, not just make vibes-based picks. Qualifying and historical-results data from OpenF1 are a big part of what it reasons over.

The games (all auto-scored off real session data)
Race predictions — podium, fastest lap & first retirement.
Grid Challenge — predict the top-10 qualifying order.
Pit Wall / Strategy — call the race winner's starting tyre, pit-stop count, first-stop window, and whether a safety car appears.
Fantasy league, trivia, and a few arcade-style mini-games.
Everything auto-scores from the data once a session finishes — no manual entry.

Stack (for the curious)
Next.js + React, PostgreSQL, a small live-WS service bridging the MQTT feed to browsers, and Bunny CDN up front. Recently moved it onto a hardened, containerised server.

Would love any feedback — especially from people who know the data well, since edge cases (wet-tyre stints, red-flag classifications, stewards' penalties) are exactly where it gets interesting. And thanks again for OpenF1 — it makes projects like this possible.

Pitbrain.ai

[br-g]

Looks very nice and comprehensive!
Curious how the race oracle is working. Is it powered by an LLM?

[GCloke]

Yeah — it's LLM-powered. The pipeline runs in two stages: first it pulls current context (recent form, standings, news, conditions) into a structured briefing, then a second pass turns that plus the season's results into a podium prediction with a per-pick confidence. It fires automatically ~30 minutes before lights-out and self-scores against the actual classification once results land.