Criteo x NVIDIA AI Meetup | LLM-Powered Recommendations at Scale

Overview Join NVIDIA and Criteo to hear how teams take LLM-powered recommendations from prototype to production. Learn what it really takes to deliver relevance at scale, from inference stack design to optimizing open-weight models on-premises, with candid lessons from the builders. Agenda 6:00 PM - Door openings 6:30 PM - Welcoming Remarks by Criteo and Nvidia 6:45 PM - How to Build a High-Throughput, Low-Latency Product Relevance Classifier On-Premises w/ Criteo 7:15 PM - Advancing Generative Recommendation Inference w/ NVIDIA 7:45 PM - Panel Discussion 8:15 PM - Networking, food and drinks Sessions Advancing Generative Recommendation Inference Marek Pikulski, Senior AI Developer Technology Engineer, NVIDIA Learn how NVIDIA and Spotify engineered a production-grade inference stack for LLM-powered personalized recommendations. This talk covers the specific challenges of running beam search, multi-item candidate scoring, and user soft token injection at scale, and how native support for these patterns was built into TensorRT LLM. We’ll walk through the architecture decisions, open-source contributions, and the benchmarks that made the case for the switch. How to Build a High-Throughput, Low-Latency Product Relevance Classifier On-Premises Zakaria Ouzarf, Software Development Engineer, Criteo Marius Alonso, Machine Learning Engineer, Criteo How can you build an open-weight LLM-based product relevance classifier that matches the performance of a frontier model—entirely on-premises? We’ll share our journey building a relevance classifier for agentic recommendations, covering fine-tuning, distillation, quantization, and inference optimization, along with the wins, dead ends, and hard numbers. Resources and Legal Resources NVIDIA Privacy Policy: https://www.nvidia.com/en-us/about-nvidia/privacy-policy/ Criteo Privacy Policy: https://www.criteo.com/privacy/