I build systems close to the machine: language models from first principles, operating-system experiments, physically inspired media tools, and browser software that fixes things the browser should already do.

My work starts with one question: what is the machinery actually doing? Then I make it visible, measurable, and changeable.

A from-scratch, byte-level language-model project with a C/C++ systems-first stack.

• Dependency-free C11 model core.
• C++17 training and inference harnesses.
• Optional LibTorch training path and Python checkpoint inference.
• Raw byte IDs [0, 255]; no pretrained tokenizer or model weights.
• A roughly 101.6M-parameter decoder trained on TinyStories, CPU-only on an M1 Max, reduced training loss from 5.72 to 1.13 over 2,000 steps and began producing recognizable story structure.

A research OS project centered on Apple's open Darwin/XNU substrate. The repository is the command center for notes, manifests, experiments, and patches—not a macOS clone or a UI skin.

A C++ video-processing project built to simulate analog signal degradation instead of pasting a generic VHS filter over RGB frames. Its current in-memory effects include chroma degradation, line jitter, dropouts, and head-switching noise.

• hover-zoom — a Safari-focused media preview userscript with hover previews, movable popouts, and copy, open, and download controls.
• insta-loader — a maintained Safari-focused userscript fork for Instagram media tools, batch downloads, and an in-page image viewer.

• Build the smallest honest version.
• Keep ownership, memory layout, and control flow visible.
• Prefer measured behavior over architectural wishcasting.
• Make interfaces boring and internals understandable.
• If something is working well and is not inefficient, do not refactor it.

C · C++ · model architecture · training and inference · byte-level language models · mixture-of-experts · Darwin/XNU · browser tooling · media systems

Based in Idaho. Interested in difficult research and engineering problems where model architecture, systems design, and product judgment meet.