High-Performance Engineering: Forward Deployed at Aston Martin F1
A session on CoreWeave's forward deployed engineering partnership with the Aston Martin Aramco F1 Team, covering why Formula 1 is an unforgiving environment for AI, and three projects: a unified data platform for aerodynamic and telemetry data, Delta to Map, a small transformer model that reduces the gap between wind tunnel predictions and track measurements, and Radio Insights, a fine-tuned speech-to-text and agentic system that makes sense of 40 live radio channels. The discussion extracts lessons applicable to any team moving AI from experiment to production under fixed deadlines.
Formula 1 is framed here as a technological arms race in which every result is public and every race is a test, and this session walks through how CoreWeave’s forward deployed engineers have worked inside the Aston Martin Aramco F1 Team for roughly 14 months, alongside the team’s own engineers. The work is organized into three projects: a data platform, Delta to Map, and Radio Insights.
The Team and Its Context
Eric Erns, the team’s commercial technology ambassador, sets the scene. The team sits in a cluster of about eight Formula 1 teams within roughly 20 kilometers of Silverstone, a dense ‘valley’ where engineers move between teams and everyone is chasing the last few percent of performance. Advantage comes from how well a team uses data, since cost caps and wind tunnel regulations limit the other levers. Aston Martin re-entered the sport under new ownership in 2021, built its factory at Silverstone in 2023, and employs about 1,100 people who build the car largely by hand. Roughly three quarters of the company looks like any other enterprise, with HR, commercial and communications functions, while the remainder is racing engineers and designers. In 2025 the team brought its own wind tunnel online, which it calls an ‘air tunnel’ in a playful aside, used for aerodynamic development and vehicle modeling.
Eric also describes the team’s technology stack as a layered set of partners, from silicon and data center providers (CoreWeave for compute, NetApp for storage) through the LLM and sovereign AI layer, security, and orchestration. The argument is that bringing in people from outside motorsport gives an engineering team focused on a single car access to the latest AI and GPU techniques.
Forward Deployed Engineering in Formula 1
Morgan Maguire, who leads the applied AI team at CoreWeave, explains why a Formula 1 team looks unlike most customers. There are no second chances: 24 races a season each count for points, and each race is also a learning opportunity for the next one. Stakes are concentrated in a roughly 72-hour race weekend, where more than 100 people, both trackside and in the factory mission control, make split-second decisions that cannot be reversed. Projects are not organized around quarterly or annual goals but around the race calendar and the date by which the car must reach a given level.
This shapes the forward deployed model. The CoreWeave team spent time sitting alongside Aston Martin staff in the factory, at the track, and during race weekends, including night and early-morning shifts that follow races across Asia, Europe and the US. Morgan’s central claim is that the most valuable projects are often the ones nobody says out loud, and they only become obvious when you watch the work happen on the ground.
Project One: The Data Platform
The first project addresses fragmented data sources. Aerodynamic simulations model tens of millions of points on the car and produce files hundreds of gigabytes in size, with thousands run per year, so storage and fast serving for later visualization and iteration are essential. Wind tunnel sensors must be ingested as well. On race day, live telemetry streams from over 100 sensors covering pressure, engine, gearbox and tire temperatures, with low latency to track side and to the UK. Bandwidth limits mean that far more data is recorded on the car than can be streamed live, and connection dropouts, such as a missing corner, must be reconciled with the complete post-race data.
The platform aims to unify these sources with role-based access control, deduplication so everyone works from the same data, and clean APIs. Morgan frames this as the foundation for replaying race data in a simulator to test setup changes and, eventually, for giving agents and models access to data sources that were previously hard to use.
Project Two: Delta to Map
Delta to Map targets the gap between the internal model’s predictions, drawn from the wind tunnel, and what the car measures on track, specifically downforce and drag derived from aerodynamic sensors. In the week before a race, the team must decide on front wing flap settings, floor choice and ride height, and wind tunnel data alone cannot perfectly represent the real world.
The problem is framed as a regression task. Linear and logistic regression served as baselines, and boosting models were also tried. The surprising result was that a small transformer, trained on a relatively small number of races, was the most effective architecture. Morgan contrasts this with LLMs that require trillions of tokens and notes that the team did not have anything close to that scale of data. The key ingredient was close collaboration with Aston Martin’s vehicle science team, whose physics intuition was translated into model features. The engagement also surfaced specific features in the team’s data that gave them deeper insight into where the discrepancy between wind tunnel and track data originates, which Morgan presents as a lasting deliverable beyond the model and training script.
Project Three: Radio Insights
Agata Murachik presents the radio problem as an intelligence advantage buried under noise. Teams can listen to competitor radio, and drivers and race engineers discuss pit strategy and tire condition, but a new audio message arrives roughly every two seconds, and there are about 20 internal channels alongside the competitor streams. No engineer can monitor around 40 streams at once while doing their own job.
The Radio Insights platform captures and transcribes all 40 channels continuously, formats the messages for quick reading, automatically categorizes them by relevance (so a tire engineer is not distracted by irrelevant traffic), and offers a chat interface for asking questions about what has happened on internal or competitor channels. During a race it supports faster decisions, and afterwards it supports strategy reviews, communication and competitor analysis.
Transcription quality is the core of the system. Speech in a race car, at around 300 kilometers per hour with heavy background noise, is far from clean, and the vocabulary includes technical terms like ‘Strat 7’, driver names such as Hulkenberg, and, as Agata notes with some amusement, a surprising amount of swearing. The team spent a race weekend learning how the radio is used, then hand-annotated seven hours of production audio, yielding 3,000 annotated samples. Reaching production quality took 75 model iterations, with Weights and Biases used for experiment tracking, dataset and model versioning, and Agenta for evaluation. The platform was tested live at the 2025 Qatar and Abu Dhabi races and has run throughout the 2026 season. During the Monaco 2026 weekend it processed more than 50 hours of audio, roughly 21,000 messages across all 40 channels.
The benefits are twofold. During a race, engineers can stop monitoring all traffic and retrieve information when they need it. After a race, the platform provides a searchable, evidence-based record for debriefs and competitor analysis, rather than relying on recollection. The structured output can also feed automated post-race debriefs and be combined with telemetry and timing data for deeper strategy work.
Takeaways
Across the three projects the common thread is that forward deployed engineering only works when it is grounded in the customer’s daily reality, when domain experts supply the physics and operational intuition that generic models lack, and when small, carefully curated datasets combined with the right architecture can beat scale. The session closes with the idea that the goal is to let engineers spend less time monitoring noise and more time making the decisions that matter.