About

I’m Kevin, a computer science student at Cornell. I like the part of machine learning where the model meets everything around it: the serving stack, what ends up in the context window, and the numbers that tell you whether a change actually helped.

This summer I was at Gail, working on the realtime voice agent that answers the phone for insurance agencies. Most of my work lived at the edges of a live call: switching speech-to-text providers mid-call so an email address gets spelled out right, racing a second LLM request when the first one stalls, filler phrases in sixteen languages so a slow tool call doesn’t sound like a dropped line, and tracing so you can see where every turn spent its time.

Right now I’m running long-lived agent loops on my own hardware (two 3090s, two Qwen models) and writing up what makes them better. On the side: whether Muon works because its Newton–Schulz step approximates the polar factor, or because of the shape it gives the singular values; a small cluster manager for local GPUs; and a demo for face recognition that runs on encrypted images.

Background

Education

B.S. Computer Science, Cornell University, 2026

Previously

Gail — realtime voice agents, summer 2026

Now

Local agent loops, optimizer research

Location

Ithaca, NY

Contact

kjr64@cornell.edu is the best way to reach me. Code is on GitHub, and I’m on LinkedIn.