Large language models have grown remarkably capable, but their scale makes them costly and slow to deploy. This project investigates methods for making LLM inference more efficient, spanning model compression, low-precision numerical representations, and hardware-aware kernel design, to preserve model quality while substantially reducing the memory and compute required to run these systems. By studying where accuracy degrades under aggressive optimization and developing techniques that adapt to the structure of a given model, the work aims to narrow the gap between state-of-the-art models and the resources needed to serve them. The broader objective is to contribute practical, generalizable approaches that make powerful models more accessible across a wider range of applications.

I’m drawn to SuperUROP because it offers the chance to take a research question seriously over a full year, to move past isolated experiments and see an idea through from first intuition to a result worth sharing. Working at the intersection of machine learning and systems, I’ve found that the most interesting problems live in the details of how models really run. I’m excited to grow as a researcher and contribute something meaningful to the field of efficient AI.