platoseedplatoseed
Reinforcement Learning (RL) for AI Agents
Osmosis is a post-training platform that helps companies fine-tune language models using reinforcement learning. We work with fast-growing AI companies to train task/domain-specific models that beat foundation models on performance, cost, and latency. Our platform handles compute orchestration, reward modeling, and training run observability as a CLI-based product usable by developers and agents.
Osmosis is a forward deployed reinforcement learning platform that collaborates with customers to post-train and fine-tune agents using RL techniques, aiming to improve task-specific models with continuous retraining and hands-on support. It emphasizes reducing infrastructure headaches and enabling rapid updates, balancing performance with customer specifications. The product targets enterprise use cases across data extraction, code generation, and multi-tool agent workflows.
Osmosis provides a post-training reinforcement learning platform that supports feature engineering, reward function creation, and hands-on model training and serving. It enables engineers to leverage RL techniques (e.g., GRPO, DAPO) and multi-turn tool training without managing underlying infrastructure. The platform integrates with evaluation solutions to monitor performance and automatically trigger retraining, ingesting real-time data to update models as frequently as hourly. Use cases include building domain-specific extraction models, teaching AI agents to use production tools, and training domain-specific code generation models for fast generation of languages, components, and context-aware tests.
Who it’s for: Enterprises and teams building task-specific AI agents and production-grade models that require ongoing post-training tuning, tool-use capabilities, and rapid retraining loops.
branding around a hands-on enterprise platform with ongoing retraining and deployment support; mentions of platform details and customer-specific deployments, suggesting active productization and potential engagement though no specific customer names present.
Defining how real time systems should be built. Previously worked on data architecture and ecosystem at TikTok recommendations as the 5th and youngest US hire
Osmosis gives AI agents the ability to learn like DeepSeek R1, all in real-time
Osmosis provides a framework for AI agents to learn in real-time from their tasks by maintaining a library of past interactions, querying for relevant scenarios, and recommending action plans. It targets AI developers seeking improved accuracy, speed, and cost efficiency, demonstrating faster performance and fewer steps than baselines in real-time learning.
Formerly “Gulp” · why startups rename →
Reinforcement learning as a service
Open-Source Reinforcement Learning (RL) & Fine-tuning for LLMs.
