In elite motorsport, strategic windows last seconds: a one-lap early pit stop can mean a podium; a late one can cost a championship. Formula 1 teams generate over 100 GB of data per race weekend, yet a race engineer's ability to synthesise competing signals in real time is fundamentally capped by human cognition. With performance margins measured in thousandths of a second, competitive advantage now lives entirely in the speed of insight.
The challenge was never simply training a model on historical data. It was building a system that could reason over heterogeneous live streams (CAN-bus, IMU sensors, tyre pressure, thermal imaging, 1 kHz GPS) with latency lower than a single radio exchange. In parallel, real-time CFD demanded aerodynamic flow simulation faithful enough to be operationally useful, yet fast enough to run mid-event rather than hours later.
Zeklar designed a three-layer architecture: a Rust ingest layer that normalises and timestamps every signal within 200 µs of arrival; a multimodal transformer inference engine trained on a proprietary corpus of 14 annotated F1 seasons, paired with a fine-tuned LLaMA 3 grounded in FIA regulations, engineer briefings, and race-radio transcripts; and a visualisation layer that renders CFD flow fields in NVIDIA Omniverse and streams them to the pit wall through a WebGL/Three.js console. Tyre degradation models combine physics-informed neural networks with live thermal feedback, achieving lap-level prediction accuracy above 94%.
Sub-2 ms Inference
The full telemetry pipeline, from raw signal receipt to synthesised voice advisory, completes in an average of 1.2 ms, well inside a race engineer's decision window.
Real-Time Generative CFD
A neural surrogate trained on 40 million offline CFD solutions reproduces full aerodynamic flow fields at 60 fps, removing traditional simulation cycles entirely during the race weekend.
Motorsport-Native LLM
Fine-tuned on over 8,000 hours of race-radio transcripts, FIA technical documents, and internal briefings, producing contextual reasoning engineers describe as indistinguishable from a human expert.
Tyre Degradation Forecasting
Hybrid PINNs forecast the degradation cliff with a mean error of 0.8 laps over a 15-lap horizon, 3× more accurate than the team's prior statistical models.
Discovery & Data Architecture
Mapped every telemetry stream, negotiated data agreements with sensor vendors, and designed the high-frequency Rust ingest schema.
Model Training & Fine-Tuning
Trained the CFD surrogate on an A100 cluster, then fine-tuned the LLM on a motorsport corpus curated alongside the constructor's own engineers.
Real-Time Integration
Wired the AI stack into live garage systems via Kafka, optimised end-to-end latency, and validated deterministic behaviour under simulated race conditions.
Deployment & Go-Live
Phased rollout across three development Grands Prix, incorporating live engineer feedback and hardening for the 2027 campaign.
Across the 2027 season, the constructor recorded its highest strategic execution rate in a decade, with three race wins directly attributed to pit windows NEXUS identified four to six laps ahead of rival models. The system is now classified as critical infrastructure for the team's 2028 title programme.