GenAI powers on-track analysis for a high-performance racing team

Yesterday’s INDYCAR Championship Final in Nashville marked the end of the 2024 racing season, and the third season of a team-up between machine learning solutions innovator Zapata AI and legendary racing organization Andretti Global.

Now, the connection between AI and INDYCAR racing may not be obvious, but when you consider that there are about 140 data sensors on each vehicle in an INDYCAR event, and that each vehicle generates about a terabyte of data per race, the connection becomes clearer.

To make the connection crystal clear, one terabyte of data takes up as much space as 500 hours of high definition video, 17,000 hours of digital music or 6.5 million electronic documents.

And all this data has to be processed in real time by the various engineers in the racing team. That doesn’t even include all the other information they get access to such as weather updates, aerial telemetry, track conditions, competitor data and simulations.

Racing decisions and AI predictions are data driven

During a Zoom call, Christopher Savoie, co-founder and CEO of Zapata AI, said his advanced quantum modeling supports many industries such as finance, military, drug development and manufacturing, but racing is truly unique.

“Racing is where — excuse the pun — where the rubber meets the road. It’s the fastest data flow with the most sensor data, IoT-type stuff that you’re really going to get everywhere,” Savoie said.

“Along with the criticality of the decision-making. It’s literally life or death on a lot of these decisions for the drivers. Right? That makes it a really interesting place to be, knowing the important role these AI models play.”

Savoie describes Zapata AI as a technology innovator that enables large enterprises and government agencies to capitalize on the promise of AI with industrial AI solutions powered by its Orquestra platform, which is designed for ultra-large-scale genAI applications with both text- and numerically based learning models. Its technical stack is depicted below.

Zapata AI and Andretti Global—Expanded Collaboration

This year, Zapata expanded his existing multi-year, multi-million dollar deal to be named Andretti’s Official AI and Quantum Partner. In the joint statement issued before the start of the racing season, CEO and Chairman of Andretti Global, Michael Andretti said the collaboration has been invaluable.

“Zapata AI’s generative AI and technical expertise have helped us unlock real-time competitive insights we didn’t think possible. We look forward to building on the foundation we’ve laid together, and we’re proud to have them as our official AI and Quantum Partner, the results we’ve seen so far have been impressive and we’re just getting started.”

Custom AI models in place are unique and the future

What makes Zapata unique is that rather than relying on a monolithic, centralized AI platform like Claude or ChatGPT – the company develops and deploys smaller groups – or ensembles – of custom AI models and innovative quantum algorithms to the places where the action happens .

Known as edge computing, this type of network design brings the digital processing and computing power physically closer to where data is generated in real time. This ensures faster data collection, analytics and on-the-spot decision making with less latency.

“So training in general is great. But when we get to these kinds of industrial use cases, what really matters is streaming live data and making sure the AI ​​model is thinking about what’s actually happening right now. That’s where these things become much more useful,” Savoie said.

Savoie explains that the custom AI models are trained on more than 20 years of past proprietary racing data from the entire Andretti team, which would be similar to pilot training in a world-class flight simulator. But with the help of Zapata’s on-site Race Analytics Command Center – powered by two Nvidia H100 GPUs – the on-site AI models receive real-time training and provide updated predictive scenarios based on the flood of data flowing from the actual race. Extending to the pilot analogy would be similar to the simulator trained pilot who flies into combat and learns under live fire with split second decision making.

These live trainings for Zapata’s custom AI ensembles further differentiate it from all other current enterprise AI offerings available in the space. Savoie believes these types of models will have a greater impact on the future than current genAI models.

“I think these smaller ensembles of specific models that are used in these kinds of boring industrial things — that’s really going to be where a lot of the behind-the-scenes work and benefits to humanity are going to happen and occur,” he said.

AI insights give Andretti global traction

Zapata has worked with Andretti Global engineers to build and deploy advanced ML models to better understand severe tire degradation analysis, identify fuel savings opportunities, lap time predictions, and provide yellow flag predictive modeling.

“So a yellow flag means there’s been an incident on the track, a car has stopped, or there’s been an accident, or there’s debris on the racetrack, or something that’s causing the other cars to slow down. They get don’t pass, and they all line up, and usually they send out a safety car,” Savoie explained.

“If you happen to be on pit row picking up fuel or changing tires when a yellow flag occurs, you’re shuffled to the back of the pack – so it’s useful to make a data-based prediction of when yellow flags are most likely to occur to avoid losing valuable race position. Basically, we could say that we predict it correctly over 80% of the time.”

During yesterday’s event, Colton Herta – a driver for Andretti Global – won the season finale. Savoie said his company’s AI solutions excel in high-stakes scenarios where immediate, accurate decision-making is critical – noting that these solutions extend beyond racing.

“This is all time series data, and we see time series data in a lot of different areas like manufacturing lines, automotive data, as well as financial modeling of market predictions and even risk modeling for insurance companies. It’s the same thing,” he said.