Hi, I’m Yash.

I build and study language models. At Wethos AI, I work on post-training, personalization, and models with persistent memory.

Most of my independent work starts with a model doing something its training objective or evaluation does not fully explain. I’ve looked at why a rank-allocation method that works under SFT fails under GRPO, whether personalization metrics measure the person or only surface resemblance, and what one RL update changes inside a model.

I tend to understand these questions by building the system, finding where the abstraction breaks, and testing the explanation directly. That is also why I build models from scratch and write about experiments that fail: both are useful ways of finding out which parts I don’t actually understand.

I came to this through applied machine learning. Before Wethos, I built language-model systems at BNY Mellon and JerseySTEM. I studied data science at UC Irvine and NMIMS University.

Resume