Leads developer relations and product for AI infrastructure and managed inference at Crusoe.
// 01 / Bio
Suman Debnath is the Director of Developer Relations and Product at Crusoe, where he focuses on AI infrastructure and managed inference. Prior to joining Crusoe, he served as a Technical Lead and Principal Developer Advocate (ML) at Anyscale and AWS, where he worked closely with developer communities to advance the adoption of machine learning at scale. With deep expertise spanning distributed systems, deep learning, and natural language processing, Suman is passionate about bridging the gap between cutting-edge AI research and practical, production-ready solutions. He is a frequent speaker at industry conferences and an active contributor to the broader ML and developer community.
// 02 / The session
Post-training at scale: engineering the RL loop for LLMs
Abstract
Reinforcement learning is now a core part of how modern LLMs are aligned, tuned for reasoning, and adapted to specific tasks. Running it at scale is a systems problem as much as a modeling one.
Each RL step chains together rollout generation, reward scoring, and policy updates, and each stage stresses the hardware differently. This session walks through the anatomy of the RL post-training loop and the engineering decisions that determine its speed and reliability.
What you’ll learn
The three stages of the RL loop
Why rollouts are an inference problem
Keeping RL fast and reliable at scale
// 03 / For attendees
Keep learning
Go deeper on what this session covers.
AI DEV · BY DEEPLEARNING.AI
Join us at AI Dev
Meet the engineers shipping AI to production, and leave ready to do it yourself.