Scratchpad
Things to remember, quick notes, rough calculations, and doodles.
-
AIVA and SAIRA: talking cartoon avatars from one home GPU
One portrait, a few video clips and five mouth pictures per character, a Piper voice, and a reverse SSH tunnel to the model at home. With demo videos.
-
Tiny Earth: one prompt, two local models, one GPU
Two sizes of Qwen3.8-Flash-Next on one 32 GB Radeon each built a working Three.js planet from a single prompt. Both planets are playable.
-
One GPU, a whole AI stack: benchmarking the Radeon AI PRO R9700
Chat, image and video models and a voice assistant on one 32 GB AMD card, with benchmarks and the ROCm bug that turned videos into noise.
-
Using Google APIs safely from a static website
Why a Google key never belongs in the browser, and how to keep it on the server with the bill at zero.
-
A homelab that costs $10 a year in power
Judge an always-on server by its idle power. What common hardware really draws, and the settings that get a whole homelab to about 7 W.
-
MQTT, Mosquitto and other brokers
The small protocol behind much of the Internet of Things, Mosquitto running in five minutes, and how the other brokers compare.
-
Big LLMs on gaming hardware with Strata
A 125-billion-parameter model on a PC with a 12 GB graphics card, and which budget upgrades are worth it.
-
Getting started with Transformers.js: AI models in the browser
Hugging Face models in a web page with no server. My Transcriber app is the worked example.
-
Compile C++ to WebAssembly: running whisper.cpp in the browser
Emscripten from install to whisper.cpp running in the browser.