Rayan Garg is co-founder and chief executive of Theta Software, which builds reinforcement-learning environments for professional work. His company trains and evaluates AI agents on demanding assignments that require navigating realistic organizational tools, data, workflows, and expert-defined standards.
Garg began building educational organizations before entering the AI industry. In 2019, he co-founded Elevate the Future with Arjun Gupta to expand access to computer science and business education. Its Project Falcon initiative taught students web development while providing free websites to small businesses during the pandemic. He also served as chief operating officer of the youth-entrepreneurship nonprofit Spark Teen.
After studying computer science at Cornell, Garg became a founding engineer at DeepSilicon, where he worked on ternary-model research. He subsequently founded Theta alongside Tanmay Sharma and Gurvir Singh. The company entered Y Combinator’s Spring 2025 cohort, initially developing a self-learning memory layer for AI agents before expanding into simulated professional environments.
- Realistic environments require connected tools and consequential decisions. Agent difficulty depends partly on how many systems an assignment spans, whether earlier actions change subsequent choices, and how much ambiguity agents must resolve independently.
- Evaluation rubrics need their own quality controls. Garg emphasizes testing scoring criteria for coverage, consistency, and agreement with subject-matter experts. Detailed reward signals help agents learn from difficult assignments instead of exploiting incomplete or unreliable evaluation.
- Financial benchmarks should reflect the breadth of finance. Existing evaluations can saturate while concentrating on spreadsheets and investment banking, leaving credit, debt, and risk underrepresented. Garg favors broader professional tasks and mean@5 evaluation, which assesses performance across multiple attempts.
His work on long-horizon agent evaluation treats realistic tasks, rigorous grading, and useful training feedback as inseparable requirements for agents expected to handle consequential professional work.
