Kasey Zhang

Kasey Zhang is co-founder and CEO of Osmosis, a platform for reinforcement learning and post-training aimed at production AI agents. With cofounder Andy Lyu he joined YC Winter 2025 and later raised a $6.3 million seed co-led by CRV and Audacious Ventures, with angels including Paul Graham and Guillermo Rauch among the syndicate named by early backers.

Gaming founder, then investor, then builder again

Zhang previously co-founded and ran a gaming startup for two years before it was acquired, then worked at venture firms focused on AI and data before returning to building. Lyu joined TikTok early on the data infrastructure team for recommendations and later became a tech lead across core app surfaces. The two have known each other for more than a decade, which they highlight in YC launch materials as the basis for a long working partnership.

Static models versus learning loops

Osmosis argues that foundation models are frozen once deployed: they can fail the same way thousands of times without updating. The product focuses on capturing real-world feedback, turning it into rewards, and fine-tuning open models so task-specific systems improve over time. On the YC company page the team frames the developer experience as a thin integration that stores prior agent interactions and recommends plans based on similar past work.

Why investors framed it as adaptive infrastructure

In a Vermilion Cliffs note on the seed, investor Ashley Smith wrote that Zhang and Lyu were "deep in the weeds of reinforcement learning infrastructure" rather than pitching hype, and described their operating loop as observe, evaluate, reward, adapt, and repeat. That investor framing matches how Osmosis presents itself: not a bigger base model, but the feedback layer that lets agents keep getting better after ship.