Shaan Desai

Shaan Desai is a machine-learning engineer and researcher focused on making enterprise AI agents reliable, observable, and safe to deploy. His work encompasses scientific machine learning, causal discovery, enterprise agent development at Cohere, and multilingual tool-using agents.

Originally from Lusaka, Zambia, Desai earned a bachelor’s degree in physics and a master’s degree in computational science at Harvard, where he researched machine-learning approaches to renewable-energy materials. Selected as a Rhodes Scholar representing Zambia in 2017, he continued to Oxford’s doctoral program in Autonomous Intelligent Machines and Systems. There, he studied physics-informed neural networks and one-shot transfer learning; an internship at QuantumBlack extended his research into causal-graph uncertainty and DYNOTEARS, a method for learning dynamic Bayesian networks.

At Cohere, Desai developed practical approaches for agents working with enterprise data, including natural-language SQL agents and financial-data agents. He also contributed to the Command A technical report, linking application-level agent design with enterprise language-model development.

Building agents that hold up in production

• Keep architecture and tool interfaces simple. Desai starts with one model and a limited set of clearly specified tools, adding specialized agents only when necessary. Straightforward arguments, concise instructions, explicit routing rules, and selective retrieval of conversation history reduce ambiguity and unnecessary failures.

• Evaluate every step, not only the answer. His approach to golden-set agent evaluation specifies expected requests, function calls, arguments, tool outputs, and final responses, making failures traceable. He matches interventions to their severity: improved prompts and tool specifications, targeted annotations, or synthetic training data and fine-tuning.

• Put humans in control of consequential actions. Desai recommends human-in-the-loop approval before operations such as sending email, with checkpoints triggered by explicit rules. His AI Engineer Summit presentation demonstrated how retrieval, application integrations, and inspectable tool execution combine in Cohere’s North platform.

More recently, Desai coauthored research on LuckyStar 111B, a Korean-English enterprise-agent model developed through collaboration between Cohere and LG CNS. The project applies his interests in dependable tool use to bilingual reasoning and practical deployment constraints.