Sharvi Tomar
engineer + researcher
I'm an engineer and an independent researcher. I'm curious about how complex systems, neural networks especially, end up representing structure internally: what shape a concept takes inside a model, and why.
Most of my research time these days goes toward mechanistic interpretability and representation geometry, though the curiosity extends well beyond any one project.
currently writing about
How transformers represent ordinal quantities, such as bracket depth, indentation level, position in a table, and numeric magnitude, and why some of these end up living on clean, curved, low-dimensional manifolds while others don't.