Local AI infrastructure, self-hosted software, home automation, and the unglamorous parts of putting infrastructure together at home and for small businesses. First person, understated, honest about what does not work.
A door-triggered morning briefing that took three evenings to debug and one line to fix. Notes on layered Cast abstractions and why the working version is shorter than the broken one.
Ternary weight models promise a step change in local inference economics. Microsoft's bitnet.cpp is the most concrete piece of infrastructure to date. A practitioner level read on where it is at and what still needs to happen.
Trying to wire a local LLM into Claude Code looks great on paper. In practice, the capability gap is still too wide for it to be a daily driver. A measured negative finding.
A small Python pipeline that summarises Slack channels overnight using a local LLM and posts the digest to Telegram. No third party AI services in the loop.
A frank write up of building a self-hosted notes app with end-to-end encryption, image attachments, and a shopping list mode. The unsexy parts took the longest.
Object detection with RF-DETR, vision-language reasoning with Gemma, speech transcription with Whisper.js, all served from a Mac Mini M4, all running without the cloud.
What works and what doesn't when you run Gemma 3 on Apple's smallest M4 box with 32GB unified memory, with practical notes on Ollama, Open WebUI, and the VRAM ceiling.