AI Engineer World's Fair 202617:01From 15% to 90% GPU Utilization: Fix the Data Pipeline, Not the ModelTarun SunkaraneniInfrastructure and deployment · Throughput optimization · Fine-tuning