Aston Martin F1 shows why embedded AI engineers beat model selection

The forward deployment argument. CoreWeave has had engineers embedded inside Aston Martin's Formula 1 operation for 14 months. The team worked race weekends, sat in mission control, and picked projects that appeared while watching operators under pressure. The talk presents three of those projects: a unified data platform, a model that maps wind tunnel predictions to track data, and a radio transcription system. It is a partnership case study, not a technical tutorial.

Why F1 is a stress test. The team operates on a 24-race calendar, with a 72-hour window from first practice to the grand prix and more than 100 people in mission control. There are no second chances and no quarterly goals. Data is heavy and inconsistent: each aerodynamic simulation can be hundreds of gigabytes, thousands run every year, and live telemetry has to be reconciled with data physically pulled off the car after a session because track bandwidth cannot carry everything.

The foundational platform. The first project is the least glamorous and the most transferable. The goal is to connect simulation, wind tunnel, live telemetry, and post-race data behind common interfaces, with access control and confidence about what was logged. The team argues that clean, consistent data makes race replay and simulation possible, and later gives agents a reliable source to work from. This is conventional platform engineering, and the story suggests it was the necessary precondition for the other two projects.

The delta-to-map win. The second project predicts the gap between wind tunnel predictions and measured track downforce and drag. The team tried statistical baselines and boosting models, but a small transformer worked best despite the lack of huge datasets. The vehicle science group supplied domain knowledge that became features, and the work also exposed a couple of data features that gave Aston Martin insight into where the error comes from. No error metrics or validation details are shared, so the claim is suggestive, not proven.

Radio insights at race scale. The third project transcribes all 40 competitor and internal radio channels continuously. A fresh message arrives roughly every two seconds, so most intelligence is buried in noise. The team collected seven hours of production audio, created 3,000 annotated samples, and went through 75 model iterations to get a transcription model that handles F1 audio. At Monaco 2026 the platform processed more than 50 hours of audio and over 21,000 messages.