Pit Wall Chaos: 12 words, real F1 radio, and a tiny LLM judge

The driver is on the radio, the tyres are gone, the gap to the car behind is shrinking, and you have about eight seconds to say something useful before the next corner eats the conversation. That's the situation real F1 race engineers live in, and it's the situation I wanted to drop a player into! Pit Wall Chaos is a small browser game where you hear a real Formula 1 team-radio clip, write one reply in 12 words or fewer, and a tiny model decides whether you saved the race or wrecked it.

Where this came from

I'd already built a dataset of F1 team radio: 14,681 audio clips across 149 Grand Prix events and 43 drivers, running from March 2018 to December 2025. The transcripts were generated with Cohere Transcribe 03-2026, and the whole thing is under CC BY 4.0. I made it because I'm interested in how people communicate fast and precisely when there's no room for a second take - race radio is a perfect case of that. The messages are clipped, full of jargon, half the words are codes for engine modes, and somehow both sides manage to understand each other.

Sitting on a dataset like that, the obvious move is to use it for something interesting. The more fun move was to make the messiness acutally playable!

The game loop

You're cast as a rookie race engineer. Each game comprises five rounds. A clip plays, the transcript shows up as a garbled radio feed, and you type one message back to the driver. The hard cap is 12 words. Tiny Aya, a 3.35B multilingual model by Cohere Labs, reads your reply and moves four meters: trust, pace, chaos, and tyre risk. After five rounds you get a race result, anything from P1 - Tactical Genius down to P13 - Strategy Department Incident or DNF - Driver Muted The Radio.

The 12-word limit is the whole design, as it forces you into the same compression a real engineer uses; you can't explain, you have to land the point.

The part that nearly broke it

I planned to pull clips straight from the Hugging Face dataset at runtime. That kept producing decks biased by year, sometimes only 2018, sometimes only 2023 or 2024. The dataset's ordering and sharding meant that "random" samples clustered by season.

The fix was to stop streaming rows and instead build a local manifest. I match transcript IDs to the F1 live-timing MP3 URLs using static files, filter clips to roughly 4 to 40 words so they're long enough to react to but short enough to read, and that leaves 12,356 playable clips. The audio still streams from the original live-timing URLs, so nothing heavy gets packaged into the Space. Startup got faster and the season variety problem went away.

The model lives in your browser

The judge is onnx-community/tiny-aya-global-ONNX, adapted from CohereLabs/tiny-aya-global, loaded client-side with Transformers.js and run on WebGPU when it's available. The Space itself is CPU-only, because it never touches the model. It serves the app, builds the deck, and redirects audio requests - that's it!

I ask Tiny Aya to return structured JSON: scores, deltas, a verdict, and a driver reaction. For a while the game kept showing fallback verdicts and I assumed the model was too small for the task, but in the end the model was actually fine - my JSON extractor was grabbing the nested delta object instead of the full judge response. Once I collected the JSON candidates and preferred the one shaped like the actual schema, the real verdicts came through. There are still local guardrails around it: word count, over-limit penalties, meter clamping, and a fallback judge if the output is unparseable. Another step that made the game actually fun was prompting the model to be chaotic, push back, and act just like a stressed driver, making the players fight hard for their points!

What it's for

This is a Thousand Token Wood entry, so the goal is to be fun and delightful, not necessarily useful. The AI is load-bearing in the most literal way, as without a model reading your reply and reacting in character, there would be no game, just a transcript and a text box. Tiny Aya is also far under the 32B cap, and the whole thing runs locally in the browser, so it lines up with the off-the-grid spirit of the hackathon. Codex did a lot of the implementation, debugging, and deployment work alongside me.

The honest caveats: the first run is a big download, the transcripts are messy and sometimes contain both driver and engineer, requiring some inventive from the player, and the judge is playful, sometimes missing actual strategy. The chaos is part of the fun!

Play it here: https://huggingface.co/spaces/build-small-hackathon/pit-wall-chaos

Michele Ciletti Pre-doctoral Researcher in Computational Linguistics, University of Foggia