Diya Girish Kumar

I study Electrical Engineering & Computer Science (EECS) and Business at UC Berkeley's M.E.T. program. My interests revolve around infrastructure and the data paths, placement decisions, and performance constraints that ultimately shape what frontier models are capable of.

I was Employee #2 at Limestone, where we raised a $3.5M preseed and built low-latency object storage optimized for GPU training and inference. I've also conducted research at SkyPilot, working on region-aware, cross-cloud scheduling.

Last summer, I was on the platforms team at Google Labs, working on infrastructure to support frontier design and coding agents.

I believe compute is the next inflection point - not just more of it, but better abstractions for how we access, schedule, and reason over it. Reshaping that layer will define the pace of progress.

And I'm just getting started.

HOME Building a Qwen2.5-1.5B Inference Engine From Scratch The Case for Naïveté and My Ongoing Experiment on Blissful Ignorance Compute as a Commodity?