Third year Woxsen BTECH. Originally from Dubai. i do ML research independently. i like swimming too.
The work is mostly mechanistic interpretability and representation geometry: I take a result that looks good and poke it till it tells me why. Usually the answer is that the model learned to notice something other than the thing we named the benchmark after.
Recent: Register, Not Risk : a paper showing that a fair amount of LLM "harm detection" is the model quietly reacting to text that sounds like an instruction manual. Mildly inconvenient for a few research communities.
Currently trying to make intuition and logic collaborate.