Mahesh Sathiamoorthy Bespoke Labs is a founder story sixteen years in the making. Long before he co-founded the company, Sathiamoorthy was a PhD student at the University of Southern California, spending five years from 2008 to 2013 researching under a professor named Alex Dimakis. Sixteen years later, in July 2026,
Sathiamoorthy and Dimakis, now reunited as co-founders, announced $40 million in combined seed and Series A funding for the AI startup they built together, betting that the biggest obstacle to useful AI agents isn’t smarter models, but better places to train them.
The Early Years
Sathiamoorthy’s academic path ran through USC’s Department of Electrical Engineering, where he completed his PhD between 2008 and 2013, working alongside professors Bhaskar Krishnamachari and Alex Dimakis, according to his own website. That advisor relationship, formed more than a decade before either would think about starting a company, turned out to be the seed of Bespoke Labs itself.
After finishing his PhD, Sathiamoorthy spent time at Symantec and General Motors before joining Google in 2014, where he would remain for roughly a decade, eventually becoming a Staff Research Engineer at Google DeepMind. There, he worked on using large language models to improve recommendation systems, including at YouTube, and is credited as the first to productionize TPUs for recommender systems at YouTube, a shift that reshaped the platform’s compute strategy. He also helped introduce Semantic IDs and generative retrieval techniques for recommender systems, approaches now deployed or actively researched at YouTube, LinkedIn, Snapchat, Alibaba, Spotify, Meta, and Kuaishou.
Finding the Problem
By the early 2020s, Sathiamoorthy had a front-row view of one of AI’s most stubborn limitations: models that could write code, answer questions, and complete short tasks impressively, but still struggled to operate reliably over hours or days the way a human employee could. Improving a model’s raw capability, he came to believe, was only half the problem. The other half was giving that model somewhere realistic to practice.
Leaving Google DeepMind in late 2023 was not an easy decision, Sathiamoorthy has said publicly, given how much he valued the work and the people there. But he left anyway, with a specific goal: to help democratize post-training, the stage after a model’s initial training where it gets refined through techniques like reinforcement learning, and where Sathiamoorthy believed the field’s biggest bottleneck was quietly hiding.
Building Bespoke Labs
Sathiamoorthy started Bespoke Labs in 2024, calling his old PhD advisor, Alex Dimakis, who joined as co-founder and chief scientist while continuing to serve as a professor at UC Berkeley. Bespoke Labs builds reinforcement learning environments and infrastructure that let frontier labs and enterprises train, evaluate, and improve long-horizon AI agents for real production use, rather than just benchmark performance.
The company’s early work centered on open research contributions, including OpenThoughts, an open reasoning dataset that has since been used to train more than 190 public models, alongside tools like Bespoke Curator for synthetic data curation and Evalchemy for evaluation and benchmarking. Rather than building simple, app-level environments through outsourced contractors, the way some competitors approach the category, Bespoke Labs builds environments that mimic real-world business settings directly, including codebases, microservices, and communication logs, so agents can be trained and evaluated against conditions that resemble actual enterprise work.
The Pitch
Sathiamoorthy has framed Bespoke Labs’ value proposition around a resource gap rather than a modeling gap. “Frontier labs, enterprises, and all organizations relying on reliable agents need access to high-quality environments,” he said in the company’s funding announcement. “This is the critical piece needed to optimize and develop agents.” Once those environments exist, Bespoke Labs helps organizations improve their agents using techniques like GEPA, a Genetic-Pareto Agent Optimizer that automates prompt and policy search to reach higher accuracy faster than manual prompt engineering.
That pitch, environments and infrastructure as the missing layer, rather than another foundation model or another thin wrapper application, is what distinguishes Bespoke Labs from much of the current AI agent funding wave.
The Round: $40 Million and Who’s In
Mahesh Sathiamoorthy Bespoke Labs’ $40 million in total funding came in two parts. A seed round of roughly $7.25 million closed in mid-2024, led by 8VC, with participation from Google DeepMind chief scientist Jeff Dean, Resolve AI CEO Spiros Xanthos, and DevRev CEO Dheeraj Pandey. A Series A of $31.75 million followed, led by Wing VC, with participation from Mayfield, The House Fund, dbt Labs CEO Tristan Handy, and individual angel investors from Anthropic, OpenAI, and Meta. Both rounds were announced together on July 6, 2026.
The company says it will use the new capital to expand its research team, scale its environment-building infrastructure, and accelerate its commercial momentum, positioning Bespoke Labs to compete for enterprise budgets as demand for reliable, production-grade AI agents accelerates.
Proof It Works
Bespoke Labs actively contributes to open benchmarks including Terminal-Bench, a widely watched test of AI agents’ ability to complete real terminal-based engineering tasks, alongside its own OpenThoughts dataset and the GEPA optimization technique. On JobBench, an agentic-task benchmark, the company’s own team has noted placing second, trailing only Fable 5, a concrete external signal of technical competitiveness rather than a purely self-reported claim.
The company has grown to roughly 40 to 48 employees since its 2024 founding, according to TFN, a meaningful scale-up for a company that only became widely known outside AI research circles with its July 2026 funding announcement.
The Backers
Mahesh Sathiamoorthy Bespoke Labs’ investor and advisor roster reads like a cross-section of the AI research and infrastructure world. Beyond its lead investors, the company counts Jeff Dean, one of the most recognizable names in AI research, and executives from Anthropic, OpenAI, and Meta among its individual backers. Its advisory board includes Tasso Argyros, VP of engineering at Databricks and former CEO of ActionIQ; Joseph Gonzalez, a UC Berkeley associate professor and creator of LMSYS and vLLM; and Greg Durrett, an associate professor at NYU.
That combination of frontier-lab researchers, infrastructure executives, and academic co-founders gives Bespoke Labs a credibility profile that is unusually dense even by AI funding’s already researcher-heavy standards in 2026.
The Competition
Mahesh Sathiamoorthy Bespoke Labs operates inside a rapidly growing AI agents market, projected by Grand View Research to expand from $10.9 billion in 2026 to $182.9 billion by 2033, a compound annual growth rate of roughly 49.6%. Within that market, Bespoke Labs sits specifically in the environments and post-training infrastructure layer, distinct from companies building agent-facing applications, AI talent marketplaces, or enterprise AI governance tools.
That positioning matters for readers tracking Denote Press’s broader enterprise AI coverage: where Runlayer focuses on AI governance and Mercor focuses on matching AI talent to work, Bespoke Labs focuses further upstream, on the training infrastructure other AI companies and labs depend on to make their own agents reliable in the first place, a genuinely distinct layer of the AI stack rather than a competing product in the same category. That distinction is likely to matter more, not less, as the broader AI agents market matures and enterprise buyers become more sophisticated about which layer of the stack actually solves their specific reliability problems.
Next Chapter
With fresh capital and a growing research team, Bespoke Labs’ near-term priorities center on scaling its environment-building infrastructure and deepening commercial relationships with both frontier AI labs and enterprises that need reliable agents for production use. The company’s continued investment in open research, through OpenThoughts, Terminal-Bench, and GEPA, suggests a strategy of building commercial credibility partly through open contributions the broader AI research community can independently verify.
Sathiamoorthy has described his goal as democratizing post-training, a framing that positions Bespoke Labs’ ambitions well beyond serving a handful of frontier labs alone, toward becoming infrastructure a much broader set of enterprises rely on as they build their own production AI agents.
What It Teaches Other Founders
The Mahesh Sathiamoorthy Bespoke Labs story offers a few clear lessons for founders considering when and how to leave a stable, high-profile role to start a company. First, reuniting with a trusted former collaborator, in this case a PhD advisor he had known for over a decade, gave Bespoke Labs a founding partnership built on genuine mutual trust rather than a fresh, untested match. Second, Sathiamoorthy’s decade of hands-on experience solving real infrastructure problems at Google, including reshaping YouTube’s own compute and modeling strategy, gave him direct, credible evidence that infrastructure-layer problems are often more valuable to solve than user-facing ones.
Third, Bespoke Labs’ emphasis on open research contributions alongside its commercial product appears to have meaningfully strengthened its fundraising story, giving investors independently verifiable proof of technical credibility beyond the founders’ resumes alone.
Frequently Asked Questions
What does Bespoke Labs do? Mahesh Sathiamoorthy Bespoke Labs builds reinforcement learning environments and infrastructure that let AI labs and enterprises train, evaluate, and improve long-horizon AI agents for production use, rather than just for benchmark performance.
Who founded Bespoke Labs? Bespoke Labs was founded in 2024 by Mahesh Sathiamoorthy, a former Google DeepMind staff research engineer, and Alex Dimakis, his former PhD advisor at USC and a current professor at UC Berkeley.
How much funding has Bespoke Labs raised? Bespoke Labs has raised $40 million total: a roughly $7.25 million seed round led by 8VC in 2024, and a $31.75 million Series A led by Wing VC, both announced together on July 6, 2026.
What is Mahesh Sathiamoorthy’s background? Sathiamoorthy earned his PhD in electrical engineering from USC, then worked at Google for roughly a decade, becoming a staff research engineer at Google DeepMind before leaving in 2023 to found Bespoke Labs.
How is Bespoke Labs different from other AI agent companies? Rather than building agent-facing applications or AI talent marketplaces, Bespoke Labs focuses on the training infrastructure layer, building realistic environments that let other companies and labs train and evaluate their own AI agents.
Conclusion
Sixteen years after Alex Dimakis first supervised his PhD research at USC, he and Mahesh Sathiamoorthy are betting $40 million that the same instinct for rigorous, foundational infrastructure work applies just as well to AI agents as it once did to their academic research together. Whether Bespoke Labs’ environments-first approach becomes the standard layer other AI companies build on, or one of several competing approaches to the same problem, will likely become clearer as its Series A capital gets put to work.
Readers interested in other founders building the infrastructure layer beneath today’s AI boom may also want to look at Puru Rastogi’s physical AI training approach at Mowito, Jay Li’s data-collection wearable at Proception, and Baran Ataş’s AI simulation environments at Talp.
