To build the autonomous machines of the future, sometimes your model needs a model.
Companies developing self-driving cars, robots that manipulate the natural environment, or autonomous manufacturing equipment collect thousands, if not millions, of hours of video data for evaluation and training.
Organizing and cataloging this video is now a job for humans, who have to watch it all. Even fast forward, this doesn’t scale. NomadicMLa startup founded by CEO Mustafa Bal and CTO Varun Krishnan, wants to solve problems for customers who have 95% of their fleet data in files.
The challenge gets tougher when you’re looking for edge cases—the most valuable data depicts events that rarely happen and can confound infinite natural AI models.
Nomadic is working to solve this problem with a platform that turns hardware into a structured, searchable dataset through a collection of vision language models. This, in turn, enables better fleet monitoring and the creation of unique datasets for reinforcement learning and faster iteration.
The company announced an $8.4 million seed round on Tuesday at a $50 million post-money valuation. The round was led by TQ Ventures, with participation from Pear VC and Jeff Dean, and will allow the company to onboard more customers and continue to improve its platform. Nomadic too won the first prize at the Nvidia GTC Pitch Contest last month.
The two founders, who met as Harvard computer science undergraduates, “kept facing the same technical challenges over and over in our jobs” at companies like Lyft and Snowflake, Bal told TechCrunch.
Techcrunch event
San Francisco, California
|
13-15 October 2026
“We’re giving people information about their own footage, whatever drives their own AVs [and] robot,” he said. “That’s what drives these autonomous system builders, not random data.”
Imagine, for example, trying to optimize an AV’s understanding that it might run a red light if directed by a police officer, or isolating whenever vehicles drive under a certain type of bridge. Nomadic’s platform enables these incidents to be identified both for compliance purposes and to feed directly into training pipelines.
Customers such as Zoox, Mitsubishi Electric, Natix Network and Zendar are already using the platform to develop smart machines. Antonio Puglielli, VP of Engineering at Zendar, said Nomadic’s tool allowed the company to scale its work much faster than the outsourcing alternative, and that its domain expertise set it apart from other competitors.
This kind of model-based automatic annotation tool is emerging as a core workflow for natural AI. Established data tagging companies like Scale, Kognic and Encord are developing AI tools to do this work, while Nvidia has released a family of open source models, Albamayowhich can be adapted to address the problem.
Varun argues that his company’s tool is more than just a tagger. is an “agentic reasoning system: you describe what it needs and it figures out how to find it,” using multiple models to understand the action taking place and put it into context. Nomadic’s backers expect the startup’s focus on this particular infrastructure to pay off.
“It’s the same reason Salesforce isn’t building its own cloud and Netflix isn’t building theirs [content distribution facilities]Schuster Tanger, a partner at TQ Ventures who led the round, told TechCrunch. “The second an autonomous vehicle company tries to build Nomadic in-house, they are distracted from what makes them win, which is the robot itself.
Tanger praises Nomadic’s talent, noting that Krishnan is an international chess player ranked as the 1,549th best player in the world. Krishnan, meanwhile, boasts that all of the company’s dozens of engineers have published scientific papers.
Now, they’re hard at work developing specific tools, such as one that understands the physics of lane changes from camera footage, or another that extracts more precise positions for a robot’s arms in a video. The next challenge, from the perspective of Nomadic and its customers, is to develop similar tools for non-optical data such as lidar sensor readings or integrating sensor data into multiple operations.
“Juggling around terabytes of video, hitting it against hundreds of models of over 100 billion parameters, and then extracting their precise insights is really insanely difficult,” Bal said.
