How Grab Personalizes and Optimizes RL Policies Using Ray

Learn how Southeast Asia's leading superapp Grab transforms real-world decision-making through a large-scale Reinforcement Learning platform powered end-to-end by Ray in this 31-minute conference talk from Ray Summit 2025. Discover how Roman Kotelnikov and Abhinav Rai from Grab tackle the unique complexity of operating across millions of users in diverse countries, each with distinct customer behaviors, regulations, infrastructure constraints, and rapidly changing real-time conditions like peak traffic, demand surges, and sudden weather changes. Explore how Grab delivers hyper-localized, adaptive digital experiences through models that are highly localized to granular real-time contexts, instantly adaptable to evolving conditions, and diverse enough for continual evaluation across numerous environments. Understand Ray's comprehensive role in powering every stage of Grab's RL workflows, from massive distributed training using Ray RLlib and Ray Tune for accelerated RL training with multi-node, multi-GPU scalability, rich algorithm suites, and automated optimization, to real-time deployment and adaptive evaluation with Ray Serve enabling concurrent model serving, contextual traffic routing, and operational scale for context-aware deployments. Gain insights into how Ray allows Grab's teams to focus on core innovation rather than infrastructure burdens, empowering them to build reproducible, modular, scalable RL systems that adapt in real time to Southeast Asia's complex and dynamic marketplace, and see how Ray unifies training, deployment, and online evaluation for large-scale RL systems powering some of the most demanding real-world applications in the industry.