Sourjya Roy

About Sourjya Roy

I was a PhD student at Purdue University. I work with Professor Anand Raghunathan and am part of the Centre for Brain Inspired Computing at Purdue. My research interests include Algorithmic optimizations for efficient ML inference/training, Large Language Models, HW/SW co-design (ML tasks), Algorithm-Architecture co-design, Efficient Computing of Neural NetworksCurrently working on algorithmic optimizations for Large Language Models.Current Works:1) Working on algorithmic optimizations for LLM (Large Language Models) focusing on auto-regressive decoder based Transformer models for reducing KV cache requirements.2) Working on Expert LLMs where there are multiple small models which are specialized for certain tasks. The user prompt is broken in a clever way to send the appropriate parts of the prompts to each individual experts to make a final cumulative decision.3) Working on algorithmic optimizations for the Softmax operation used in the computation of Attention in Transformer models.Google Scholar: & hl=en

Experience

1. Senior Software Engineer Ai Ml Frameworks

   AMD

   Dec 2025 — Present · San Jose, CA, US

Education

• Purdue University

  Bachelor of Science - BS, Electrical and Computer Engineering

• Purdue University

  Doctor of Philosophy - PhD, Electrical and Computer Engineering