Centralized feature repository for recommender system training and inference reuse

Instead of every data team rebuilding the same customer and product signals, the company keeps them in one shared library that both training jobs and live recommendation systems can use.