Projects
Three highlights. Everything else lives on GitHub.
Running tinygrad on the NVIDIA Jetson: talking to the GPU without CUDA
From tinygrad's computation graph to PTX to a cubin, then straight to the GPU through raw ioctls (GPFIFO, QMDs, MMIO) instead of CUDA. Also: standing it up on NixOS/JetPack, a drone dead-reckoning demo (IMU → temporal CNN + EKF), benchmarks against PyTorch and TensorRT, and porting tinygrad's NV backend to Tegra.
-
01
tinygrad on Jetson AGX Orin
GPU inference on NVIDIA Jetson using tinygrad's NV backend: raw GPU command queues, no CUDA runtime. Includes a Tegra port of tinygrad, benchmarks and a reproducible NixOS dev environment.
image: NVIDIA -
02
Nand2Tetris
A whole computer built from a single NAND gate: logic gates, ALU, CPU and memory in HDL, then an assembler, VM translator and a compiler for a small object-based language.
-
03
LeNet hackathon · 2nd place
Lead software engineer on a hackathon team at ASU, building a LeNet convolutional neural network in Keras.