Pinterest's ML Platform team shares a decade-long journey of building unified AI infrastructure, evolving from fragmented team-specific stacks to a cohesive platform serving hundreds of millions of inferences per second. The retrospective covers five distinct eras, highlighting how organizational alignment, executive sponsorship, and technical maturity must converge for platform adoption. Key lessons include the importance of bottom-up layered foundations, the temporary nature of any infrastructure solution, and how efficiency, velocity, and enablement multiply when modeling and platform teams work in tandem. The article details specific technical solutions like Linchpin DSL, Scorpion inference service, MLEnv standardization, GPU-accelerated serving, and the evolution toward foundation ranking models and generative AI.
Table of contents
Individual Team Stacks (2014–2015) and Early Unification (2016–2017)ML Platform’s Scrappy Era (2018–2019)Transition Years (2019–2020)Broader Alignment and Standardization (2021–2022)Get Pinterest Engineering’s stories in your inboxScaling the Frontier (2022–Present)Closing ReflectionsAcknowledgementsSort: