Everything in this logstarted rough —half-built, buggy,or flat-out broken.Now it’s working.working.

The portfolio before this one — hand-built WebGL all the way down. I wrote the shaders myself to learn what a GPU actually does. This site is its successor: quieter on the outside, same obsession underneath.

Frosted glass is everywhere on Linux desktops like Hyprland, and I wanted to know how that blur actually gets drawn. So instead of stacking someone else's filters, I wrote the pipeline myself: a custom WGSL shader and its own renderer, on wgpu.

Match history goes in from the Riot API; tilt and MMR analysis comes out — a read on whether you're actually improving or just queueing angry. Built in TypeScript, because the fastest way to learn a data pipeline is to point one at a game you care about.

An MCP server that gives an AI agent real hands on hardware. It talks to an Arduino Uno over a serial cable to list ports, read sensor data, send commands, and even analyze thrust-test runs — the bridge between software agents and the physical world.

Tooling that turns a rough idea into a drawable manga page plan — it lays out the panels and writes the image prompt for each one. A local Codex plugin of skills, built to get from concept to a page plan fast.

A Windows desktop app for capturing gameplay that keeps everything on your own machine. Record, trim, stitch, and export through an ffmpeg-backed pipeline, with a library to review your clips — all in Electron. I wanted to record without the FPS hit or the cloud.

Firmware for a virtual pet living on an ESP32-S3. Stats decay in real time, persistent storage means it remembers you across power-offs, a TFT screen gives it a face, and a local AI-dialogue layer gives it something to say. A Tamagotchi, basically — except I wrote the firmware, and a crash means the pet dies.

The full loop — training, evaluation, ONNX export, inference — so the result isn't a notebook experiment, it's a model something could actually run. Built to learn what computer vision does when the input isn't a clean daylight photo.

Most knowledge-graph views look like server diagrams; I wanted one that looks alive. The garden grows from your tags and your writing history, so old, well-tended corners of your notes read differently from what you planted last week.