read as: html · markdown

Yashwant Bezawada

Forward Deployed Engineer at FedEx. Tinkerer.

About

I keep models answering questions in production. Upstream, I once fixed a race between KV-cache loads and CUDA-graph replay in vLLM, and I still bring it up unprompted.

side projects

Built for fun and learning.

Senior Tinkerer · 2025 – now

  • snapwire.io Sends a file from one screen to another device's camera. No network.

  • morp.ai Keeps a codebase patched and current without changing what it does. Not open yet.

  • remixpdf.com A PDF editor that runs in the browser; files never upload.

  • devtoolkit.io Developer tools that run in the browser.

Tinkerer · 2024

  • removebackgrounds.ai Background removal that keeps the full-size original in your browser.

  • planwright.ai Planning for software that agents help write. Not open yet.

work

  • – now

    · Software Engineer Advisor: working as a forward deployed engineer

  • –

    · Software Engineer II: pricing systems and cloud migration

  • –

    Earlier · Navy Federal Credit Union (software engineering intern), George Mason University (graduate teaching assistant), SmatSocial (software engineer)

open source

A few fixes that got merged.

  1. vLLM · KV-cache loads could race CUDA-graph replay · #31057
  2. ROCm hipDNN · samples now fail when their output is wrong · #3552
  3. QuantLib · a dangling-reference segfault from the Python bindings · #2410
  4. JAX · jnp.arange compile time grew with the size of the range · #33602

More on GitHub

education

  • M.S. Computer Science · George Mason University

  • B.Tech Computer Science · KL University

elsewhere

me@yashwantbezawada.com · GitHub · LinkedIn · George Mason