A home research lab. Local models, local data, no vendors and no permission. Most of it runs on hardware I already had.
I build things to find out how they work. Every so often one of them turns out to be useful to somebody else.
Identity and cognition turn out to be separable. One layer says who; a second says how. Split them and the persona stops fracturing when the input gets hostile or vague.
Tested across model families and sizes. Small models mostly. Still being wrong in interesting ways.
Self-hosted orchestration doing unglamorous work: phone triage, document pipelines, alerts sorted by how much they actually matter. Some things wake you up. Most things wait.
Running on real operations, not demos. It has been wrong twice. Both were instructive.
Vision models instructed cognitively rather than behaviourally — told how to look, not what to output. Mixed pages: tables, diagrams, handwriting, whatever fell in the scanner.
Reading strategies borrowed straight from the persona work. Handwriting is still the hard part.
Several purpose-built personas take the same live data down completely different reasoning paths, each grounded in tools it has to actually call. Then I read where they split.
Consensus is boring. The disagreements are the finding.
The interesting problems sit right at the edge of what's supposed to be possible with the hardware you already own. Constraints don't limit the work — they point it somewhere.
Data that stays put is data you control. Local-first is a design principle here, not a budget decision.
Automation without a vendor still runs when the vendor doesn't. Same for the model, and the internet.
Nothing here is trying to disrupt anything. It's a lab, in a house, in Turlock.
Something is being sourced. Parts are on order, the chassis is drawn, and it hasn't earned a name on this page yet.