You are building shared ML infrastructure for a university research environment. Multiple teams train models on overlapping institutional and research data, and they want a common place to define, discover, and serve reusable features without each project rebuilding the same pipelines.
How would you design a feature store for reusable research features?