Mahesh Sathiamoorthy

Mahesh Sathiamoorthy is co-founder and chief executive of Bespoke Labs, which builds the training data and environments that help autonomous AI agents perform reliably. Previously an engineer and researcher at Google DeepMind, he helped develop generative retrieval for recommender systems before focusing on the practical infrastructure required to train agents for sustained, verifiable work.

Sathiamoorthy studied at the Indian Institute of Technology Kharagpur and earned a master’s degree and doctorate at the University of Southern California. His early research explored erasure codes for distributed storage, and his HadoopUSC implementation incorporated multiple techniques for protecting distributed data against failures.

At Google, he worked across Google Brain, Google DeepMind, and YouTube recommendation systems. He co-authored the TIGER recommendation framework, which represents items with structured semantic identifiers and treats recommendation as a generative retrieval problem.

In 2024, he founded Bespoke Labs with Alexandros Dimakis, a collaborator from his graduate-school research. Sathiamoorthy became chief executive and Dimakis chief scientist. Bespoke expanded from model customization into open reasoning datasets and agent-training infrastructure, announcing $40 million in seed and Series A funding in 2026.

What makes agents dependable

• Training data requires experimental design. Sathiamoorthy’s Bespoke Curator helps generate synthetic training examples and run post-training workflows. He evaluates data quality through concrete choices about question selection, difficulty, teacher models, answer generation, and measurable downstream results.

• Open reasoning data should expose its methods. Bespoke’s Stratos initiative developed into OpenThoughts, a collaboration with researchers at Stanford, Berkeley, and the University of Washington. Its published research, which Sathiamoorthy co-authored, documents controlled experiments showing that multiple answers to one question can improve reasoning diversity and that the strongest model is not always the best teacher.

• Reinforcement-learning environments are part of the training data. Sathiamoorthy emphasizes realistic environments for evaluating and training agents, including versioned tasks, sandboxed execution, checkpoints, generated trajectories, and verifiable outcomes. His work on agent post-training also highlights the value of supervised fine-tuning before more computationally intensive reinforcement learning.

• Production reliability includes compliance and factual precision. In work involving Intuit’s Credit Karma, structured tags helped prevent a customized recommendation model from inventing financial details while improving compliance, latency, and throughput.