tinkering

My wife and a friend ran an online pilates program with 400+ classes trapped on an expensive instructor platform. I built them their own: bulk-downloaded and transcribed the lot, used local AI to generate titles and categories, tagged instructors with local facial recognition, and stood up a full site with subscriptions, streaming, and an admin panel. Took the running cost from about £100/month to about £7. Their content, their platform, their clients still have access.

A structured, multi-agent workflow for building software with AI that produces something maintainable rather than disposable. A team of agents (planner, architect, developer, reviewer, tester) with me as the PM, plus guardrails - hooks and scripts that enforce the framework as the agents run, so the audit trail comes out of the process itself. Built, rebuilt, and still evolving. I’ve written about the journey (see ramblings).

Home Assistant runs the flat - lights, heating, blinds, media. On top of that I built a custom integration that adds a personality and persistent memory using a local model via Ollama, with a frontier model for the harder reasoning. The goal was a home that pays attention, not a speaker you shout commands at. Still a work in progress. Still a lot of fun.

I got tired of missing deliveries, so I set out to make my doorbell hold a conversation with whoever’s at the door - hear them, generate a response, speak back, all locally. Getting audio out of the doorbell meant reverse-engineering its proprietary protocol (with a lot of help). It works now. Is it practical? Questionable. Was it fun? Absolutely.

A small agent that watches specific email threads and, when the other side goes quiet, drafts a chase in my voice and drops it in my drafts for review. Runs entirely on local hardware - no email content leaves my network. Built to annoy my building management company into fixing a leak. It’s helping.