ML Researcher — San Francisco
I am a Machine Learning researcher building language models that can deconstruct reasoning problems, improvise, and critique themselves. Currently designing an effective LLM routing system at Ema.
Glacier 3000 • Switzerland
Work
Distill a large model's reasoning into small cooperating models that can learn to solve complex reasoning tasks nearly as well as much larger LMs but with far fewer parameters.
Paper ↗Self guided distillation where a student decides when to explore new learning with teacher and when to exploit and solidify what it knows.
Paper ↗A refinement paradigm where an asker asks questions to decided when an LLM should refine its outputs and a truster decides whether to trust the refined answer.
Paper ↗Socratic style decomposition to guide language models to break complex problems into simpler ones and then solve it iteratively.
Paper ↗Journey log
Last decade, I got a chance to work on some large scale projects:
Did research on distillation, refinement, and decomposition with Prof. Mrinmaya Sachan (ETH) and Nicholas Monath (Meta Super Intelligence).
Worked with Asli Celikyilmaz & Jason Weston at FAIR, Meta, with Patrick Xia & Jason Eisner at Microsoft Research, and at Alexa AI during my masters.
Worked with Marcus Liwicki (RPTU) and Seiichi Uchida (Kyushu) during my Masters. I was also the founder of NeuralSpace, a no-code platform for multilingual AI.
Thoughtful emails, spicy prototypes, student collabs, event invites—send them all. I answer fastest when you include context and write yourself.
Email ↗