Osmosis

Executive Summary

Osmosis is a seed-stage YC W25 company building a reinforcement learning fine-tuning platform with a genuinely differentiated forward-deployed delivery model — co-designing reward functions and running continuous re-training loops alongside enterprise customers building AI agents. The market timing is excellent: DeepSeek-R1's emergence in early 2025 validated the RL-for-task-specialization thesis in a very public way, and the AI agents market is growing at 46%+ CAGR. The founding team is credible — Andy Lyu's TikTok model-serving background is nearly ideal, and Kasey's prior exit and VC experience add useful complementary muscle. The single biggest risk is platform-level displacement: OpenAI already launched a competing RFT API in December 2024, Predibase (the closest direct competitor) is reportedly acquired, and every major hyperscaler is building managed fine-tuning into their core offerings — if "good enough" RL fine-tuning becomes a feature rather than a product, Osmosis needs to have locked in deep workflow dependencies before that window closes.