Christopher Savoie talks at a fast clip--perhaps a byproduct of collecting and analyzing data from Formula 1 race cars screaming around city streets and tracks at upwards of 240 mph.
Or maybe he is trying to keep up with the breakneck growth of artificial intelligence and Gen AI behind so much of computing today.
As founder and CEO of Zapata AI, Savoie is one of the class of new savvy executives versed in ways generative AI will matter in the industrial space. A spin-off from Harvard University in 2017, the publicly-traded company of 85 workers is heavily focused on advanced linear math principles to make predictions for everything from racecars on various tracks to armies on battlefields and financial analysts in international markets where inputs from sensors and other nodes bombard the world’s fastest processors. (The company name derives from Emiliano Zapata Salazar, a leader in the Mexican Revolution.)
The work Zatapa AI is perfecting uses advanced math to take sensors data to processors to build models and update those models live, he told a keynote crowd at Sensors Converge 2024 in Santa Clara, Calif.
“Gen AI is more than just large language models,” he said. “We weren’t really interested in language. We wanted to do the boring mathematical stuff. Instead, we use complex operational data, mostly numerical, digital and sounds.”
The company works on the premise of a model of models, not one massive model. “It’s not going to be one model that does it all,” he said. “We don’t want engineering doing the CFO’s job.”
Even so, the models are still monstrous in size and have been tamed, tuned, picked-over, for racing with Andretti Global teams using two ruggedized A100 GPUs that are crated around in the back of a truck from race to race, sometimes to monitor up to six Andretti cars at once.
With high volume data from the cars streaming over a 10hz wireless channel “sometimes we have to deal with incomplete or missing data.”
In one example of how Zapata AI works at race tracks, track and car conditions during a race can be used to precisely predict when a yellow flag will be posted after an accident. That information can be used to give a driver insights on when to make a pitstop for new tires to beat out an opponent into the pit before the yellow flag is flown. With the yellow flag, the pit could be closed, keeping other cars from changing tires.
Such split second decisions can apply to financial trades as well or even in clinical treatments with new pharmaceuticals.
Some gaps in sensors data can also be filled to help show when a tire is degrading and where it is degrading, he told the Sensors Converge audience. (Some companies call this synthetic data.)“We can deep fake what the car is doing by physically modeling with generative modeling. You can deep fake a human and you can also deep fake other things—a cat, a novel. Virtual sensors can deep fake a tire slipping angle.”
Virtual sensors matter because cameras and GPS are not fast enough to provide the data to predict a tire slipping angle as cars make repeated turns at high speeds.
In similar fashion to race car tires, drugs in testing can be deepfaked. The approach can help predict when nodes in a power grid will fail, which helps operation managers know more precisely where to send emergency crews in a widespread storm. Same applies to traffic control. Or crowd control.
In another example, he said stadium operations managers during a crowded event could quickly learn the best route for egress for a first aid crew transporting a person with a heart attack to the hospital.
“It will be ensembles of models,” Savoie said.
In an interview with Fierce Electronics, Savoie elaborated. “There’s a real opportunity with Gen AI combined with physical models in digital twins. There’s a great opportunity with AI at the edge with digital twinning for high fidelity, data driven chemicals and bio.
“We want to solve practical problems. These are hard, vexing problems, of how to make it so that things are more manufacturable, in the bio industry for example.”
In bio labs, inserting a sensor into a tank will change the results, so a virtual sensors can examine the protein level in a living cell.
Savoie’s description of Zapata AI’s work is not just fast talk. The company pulled in $5.7 million in revenues last year and has signed a number of deals that show the range of applications.(However, Zapata AI stock is facing challenges, trading on the Nasdaq in the pennies, at 0.57 on June 28 and down from 13.60 on March 28 ,four days before official trading began.)
Most recently, Zapata AI worked with KPMG UK to streamline insurance compliance models. Zapata engineers and KPMG optimized tens of million of inputs in a compliance model, reducing compute time by 1,000 times, according to a press release.
In a separate announcement, Zapata AI joined the KT Consortium, a body of chemical and bio manufacturing companies to collaborate on virtual sensors and predictive modeling with Gen AI. The consortium’s members include Henkel, Mitsubishi Chemical, Syngenta and Arkema.
Also, Zapata AI published findings from its work in Phase II of DARPA’s Quantum Benchmarking program that the company claimed offer insights into the economic value of quantum computing.
In addition to Andretti Global, Zapata AI includes as customers BBVA, BP, BASF and BMW.
Savoie summarized Zapata AI’s work for Fierce, even as he was getting mentally prepared for a Toronto F1 exhibition in coming days.
“Our ability is to create purpose-fit models that run on the cloud or the edge. We are firmly embedded in… sensors combined with AI to take things to a different level in industrial.”
Related: Zapata AI, Andretti race into deeper generative AI