C:\WINDOWS\Desktop\nick\index.html_X
*** WELCOME TO MY HOMEPAGE *** you are visitor number 000114 *** my element of choice is CUDA, atomic number: divisible by 32 *** this page is best experienced with the sound ON ***

Hello and welcome to my web page!

My name is Nicholas Patapoff and I make graphics cards do things they were not strictly designed to do. I write CUDA kernels, build GPU-accelerated ML pipelines, and every so often I wander down to the bare metal and write firmware for microcontrollers.

I started out as a chemist — a real one, with a lab coat and a B.S. in Chemistry — before discovering that the thing I actually liked about chemistry was the simulation part. These days I care about the same things I always did: what happens at the smallest scale, and how fast it can be made to happen. Memory hierarchy, kernel occupancy, and shaving milliseconds off an inference loop.

I live in San Jose, California. If you have a problem that is embarrassingly parallel, we should talk.


Cool Stuff I Made

  • A100 MatMul Optimizer — hand-tuned CUDA matrix multiply using shared memory tiling and warp-level primitives. NEW!
  • ProteinSimGPU — molecular dynamics engine for proteins, with both CPU and CUDA backends. My chemistry degree, finally justified.
  • Modal A100 Playground — serverless GPU experiments: distributed training, custom ops, multi-GPU communication patterns.

Where I've Worked

Hardware Engineer — PurcellAI (2025–now)

  • Built a full-stack data pipeline, from serial capture on embedded devices up to orchestration in Python.
  • Firmware for the Arduino Nicla Vision (Cortex-M7) in C++ and Mbed OS, including hand-rolled SPI routines.
  • Deployed quantized models on-device: 90% accuracy at 120 ms real-time inference.

Computational Physics Engineer — LongShot Space (2024–2025)

  • CUDA-accelerated 1D simulation of a booster-aided light gas gun — 50× faster than the CPU version.
  • Pipelines running 500+ concurrent simulations on a single GPU, extended to multi-GPU.
  • Profiled and tuned kernels for ~30% better memory throughput under large-batch loads.

Characterization Engineer — DragonFly Energy (2022)

  • Used FTIR, RAMAN and QCL spectroscopy to design new methods for battery degradation analysis.
  • Chased down failure mechanisms: dendrites, delamination, gassing, and bad SEI formation.

School

  • M.S. Computer Science — Arizona State University, 2025
  • B.S. Chemistry — University of Nevada, Reno, 2022

Sign my guestbook! (it is just my email, sorry)
nepatapoff@gmail.com
GitHub  |  LinkedIn  |  Resume (PDF)

*~* Visitors *~*

000114
people have been here!

About Me

  • Location: San Jose, CA
  • Occupation: GPU / hardware engineer
  • Favorite API: CUDA
  • Favorite number: 32 (warp size)
  • Favorite element: also CUDA
  • Status: caffeinated

My Toolbox

  • CUDA / C++ / C
  • PyTorch, Triton, RAPIDS
  • Multi-GPU & profiling
  • Mbed OS, SPI/I2C
  • Docker, K8s, Ray
  • Python & SQL
🚧
UNDER
CONSTRUCTION
(perpetually)
BEST VIEWED WITH
NETSCAPE 4.0
MADE WITH
NOTEPAD
VALID
HTML 4.01
POWERED BY
CUDA
800x600
OR BUST
Done28.8 kbpsInternet zone