Very cool work.
Always seemed weird to me that we dump wildly unfiltered data into models in random order

Be using a separate smaller data selection model this team reduces training burden for same performance by 50%