Pinterest's AI platform has evolved significantly over the past decade, marked by distinct phases that reflect the company's journey in machine learning. The early years (2014-2017) were characterized by fragmentation, with different teams building isolated solutions using various machine learning frameworks like scikit-learn and xgboost. This lack of standardization led to problems like training-serving skew, which negatively impacted model performance. In response, Pinterest initiated efforts toward unification, although these were often slow due to competing priorities and a lack of organizational incentives. From 2018 to 2019, a small two-engineer ML Platform team emerged amid larger product teams. Their initial goal was to demonstrate enough value to secure further investment. They quickly learned that addressing the loudest complaints didn't always lead to widespread adoption. The introduction of EzFlow aimed to improve iteration velocity by simplifying the orchestration of training workflows, but its adoption was sluggish as teams were reluctant to shift from familiar systems without immediate financial benefits. Significant projects during this time included the development of the Training Compute Platform and the introduction of PySpark, which later gained traction within a dedicated Data Engineering team. Another pivotal initiative was Galaxy, a unified signal platform that transformed complex Hadoop jobs into manageable modular signals, laying the groundwork for a unified ML feature store. These early investments, while slow to yield results, eventually created a robust foundation for Pinterest's machine learning capabilities.