From OOMs to 2x Throughput: Scaling Multimodal Data Curation on Ray | Zoox | Ray Summit 2026

Curating pre-training data for multimodal driving foundation models means extracting camera, lidar, and vehicle-state data from terabytes of sensor logs, where underlying C++ libraries can cause memory leaks, hangs, or segfaults that are hard to isolate. At Ray Summit 2026, Anindya Saha, ML Platform Engineer at Zoox, shows how combining Ray Data streaming with Ray Core task isolation curates more than 7 million samples per run across 200 nodes, at over 2x the efficiency of the previous pipeline. You'll leave with the architecture, failure-isolation strategy, and structured observability that made these runs scalable and recoverable. Liked this video? Check out other Ray Summit breakout session recordings: https://www.youtube.com/playlist?list=PLZNaJyYZRPdo Subscribe to our YouTube channel to stay up-to-date on the future of AI: https://www.youtube.com/c/anyscale 🔗 Connect with us: LinkedIn: https://www.linkedin.com/company/joinanyscale/ X: https://x.com/anyscalecompute Website: https://www.anyscale.com/