The most revealing sentence in Mahesh Sathiamoorthy's story arrived at the end of a restrained LinkedIn update. In late 2023, after roughly a decade at Google, he announced that he had left what he called his dream job at Google DeepMind. He thanked colleagues, managers and mentors. He admitted the decision was difficult. Then he offered a compact plan: “Once that's done, it's time to build!”
What he built was Bespoke Labs, an applied AI research company in Mountain View. Sathiamoorthy started it in 2024 with Alex Dimakis, the researcher who had co-advised his doctoral work at the University of Southern California. Their company sits in a part of artificial intelligence that rarely produces the prettiest demo. It works on the practice field: data curation, reinforcement-learning environments, benchmarks and the machinery that tells an AI agent whether it has actually completed a task.
That choice makes more sense when his career is viewed from the bottom up. Sathiamoorthy has repeatedly worked on the layer beneath the obvious product. As a graduate student, he asked how data should be stored when machines fail or network connections come and go. At Google, he worked on the compute and modeling systems beneath YouTube recommendations. At Bespoke, he is working on the worlds beneath AI-agent behavior.
The road begins with broken machines
Sathiamoorthy earned a B.Tech with honours from IIT Kharagpur, then moved to USC for an M.S. and Ph.D. from 2008 to 2013. His doctoral research concerned coding for distributed storage systems. The practical problem was easy to state: cloud data lives across many machines, machines fail, and the system needs an efficient way to recover without keeping wasteful copies of everything.
His research reached beyond conventional data centers. One project examined content stored across vehicles in a network, using real taxi traces to compare coding with simple replication. Another became “XORing Elephants,” a 2013 VLDB paper on locally repairable codes for Hadoop. The title was playful; the engineering problem was not. Repairing one missing piece of encoded data could otherwise require contacting many machines and moving far more information than the missing piece contained.
This was an education in constraints. Reliability was not an abstract score. It depended on bandwidth, storage cost, repair time and the behavior of a whole system. More than a decade later, Bespoke's language about AI agents sounds different, but the instinct is familiar: inspect the environment, define failure precisely and design the surrounding system to recover.
Teaching YouTube what comes next
After finishing at USC, Sathiamoorthy joined Google. His public account of the period describes work spanning Google Brain, DeepMind and YouTube. He says he was the first to productionize Tensor Processing Units for recommender systems at YouTube and helped demonstrate the gains from scaling recommendation models. Both claims point to a job where research had to survive contact with an enormous production system.
In 2019, he co-authored a paper describing YouTube's large-scale multitask ranking system for the question hiding behind every “Up next” tile: what should someone watch after the current video? The system had to balance multiple, sometimes competing objectives. It also had to account for selection bias. Academic neatness met the messy sequence of human attention.
His later work pushed recommender systems toward generative models. “Recommender Systems with Generative Retrieval,” published at NeurIPS in 2023, represented items with semantic identifiers and trained a model to generate those identifiers. Instead of treating every candidate as a flat entry in an enormous catalogue, the model could use a learned vocabulary of meaning. Sathiamoorthy says this line of work has been deployed or studied across services including YouTube, LinkedIn, Snapchat, Alibaba, Spotify, Meta and Kuaishou.
The Google years also produced visible markers of recognition. On his site, he lists more than 50 peer and spot bonuses, more than five performance awards, a best-paper award and a Feats of Engineering award. A 2023 paper he co-authored on training stability for multitask ranking models won the KDD Data Science Track best-paper award. Yet his departure note did not read like a victory lap. It read like an engineer closing one loop before opening another.
A place for agents to make mistakes
Bespoke began with data curation. Its open-source Curator library helps teams create and manage synthetic-data pipelines, with support for batch inference, structured outputs, caching and retries. OpenThoughts, which Sathiamoorthy says Bespoke was instrumental in starting, collects open reasoning data and recipes. MiniCheck focused on factuality; MiniChart handled chart question-answering. The individual projects differ, but each makes model improvement more inspectable.
Then the industry moved from chatbots toward agents that operate software. The unit of work grew longer. A useful agent might open a terminal, inspect a codebase, edit files, run a test, notice the failure and try again. A polished paragraph is no longer enough. Training needs realistic tasks. Evaluation needs a verifier that can look at the state of the system and decide whether the assignment is finished.
This is where reinforcement-learning environments enter Sathiamoorthy's vocabulary. An environment is both classroom and exam. It gives an agent tools and consequences. It exposes mistakes that a simple text comparison would miss. On the TWIML AI Podcast in 2025, he discussed reinforcement learning as an alternative to repeatedly adjusting prompts, especially for multi-step tool use. Data curation, error analysis and reward design were not supporting details. They were the work.
Terminal-Bench turns that idea into a benchmark built around command-line tasks. Bespoke describes itself as a core contributor. Its current site also highlights GEPA, an optimizer that searches prompts and policies using feedback. Together, the projects form a loop: curate better experiences, test behavior in a realistic setting, study the errors and improve the policy.
The collaborator returns
Bespoke also reunites Sathiamoorthy with Dimakis. Their connection dates to the USC years, when Dimakis co-advised Sathiamoorthy's Ph.D. and collaborated on distributed-storage research. Dimakis later became a professor at UC Berkeley. At Bespoke, he serves as co-founder and chief science officer while Sathiamoorthy serves as CEO.
That pairing helps explain the company's research-first cadence. Bespoke publishes papers, releases datasets and libraries, contributes to shared benchmarks and sells infrastructure to frontier labs and enterprises. Its advisers include researchers Joseph Gonzalez and Greg Durrett, along with Databricks engineering executive Tasso Argyros. The company says more than 200 teams use GEPA in production and OpenThoughts receives more than 10,000 monthly downloads on Hugging Face.
In July 2026, Bespoke announced $40 million across previously unannounced seed and Series A rounds. Wing Venture Capital led the Series A, with participation from Mayfield, 8VC and The House Fund. Sathiamoorthy's explanation for the delayed seed announcement was revealing: the team had become busy building MiniCheck, Curator, OpenThoughts and Terminal-Bench. It was a founder's version of missing the press release because the code shipped first.
The eye behind the systems
There is a smaller, more personal artifact in Sathiamoorthy's archive. In 2011, while at USC, he contributed photography to Spark magazine. He wrote that two years earlier he had disliked friends who spent trips behind a camera. Then he acquired a DSLR and gave it a shot. The magazine's editors described him as a regular contributor with a knack for placing ordinary objects in interesting situations. His anniversary portfolio explored architecture in monochrome and creativity through seeing things in a different light.
“I don't really like to limit myself to a specific set of interests,” he wrote, “because I try to try my hand at everything!” It is the rare early quote that still fits the later résumé. Storage codes, ranking systems, generative retrieval and reinforcement learning are distinct fields. Photography is farther away still. The continuity lies in framing: decide what matters, remove distraction and make structure visible.
Bespoke's ambition is now longer-horizon. Sathiamoorthy has written about a future in which agents can operate autonomously for weeks or months. Reaching that future requires a less romantic kind of patience. A training environment must be detailed enough to reflect production. A verifier must know the difference between looking finished and being finished. A dataset must contain hard, useful examples rather than merely more examples.
The visible agent will get the applause. Sathiamoorthy is building the place where it rehearses.