Ah apparently it's actually a @bfl_ml model, not an inhouse one — that explains why it seems to be so damn good.

→ View original post on X — @jeremyphoward