Aadi Malaviya

The work I keep coming back to is making opaque things inspectable. In UCLA's NLP lab that was literal: finding which layers and attention heads carry a behaviour, then steering them to prove the link was causal and not coincidental. At Ryniant it meant a synthesis route a chemist could trace to a citation instead of taking on faith. At Chelle it meant a rubric, so a prompt decision became something you could run an experiment on rather than argue about. At Genter it meant a required distribution a client could actually be shown, deadline attached, instead of a number living in someone's spreadsheet.

A system you cannot inspect is one you can only trust or abandon. Most of what I have built exists to give people a third option.