Scoby: I built a free Kimchi
Kimchi’s shape (plan, build, judge, three models) on free inference APIs. What free actually means, why compaction had to be turned inside out, and why the cheapest checkpoint is still a person reading the plan.
Kimchi’s shape (plan, build, judge, three models) on free inference APIs. What free actually means, why compaction had to be turned inside out, and why the cheapest checkpoint is still a person reading the plan.

A managed service is a deal: you pay to not think about what’s underneath. I self-hosted Supabase on a 16GB box over a weekend, with AI help, and made every component earn its place. What I bought back was understanding.

Same builder model, same task, two harness modes — a real feature on my homelab, and why the architecture, not a smarter model, decided every outcome that mattered.

I gave a second AI read access to the Git repos that define my k3s homelab. The findings were fine — the work was validating each: which to accept, which to recalibrate, and which the scanner was confidently wrong about.

Thirty posts, no covers, and I can’t draw. The fix wasn’t a better prompt — it was picking the right model tier and letting the model read the whole post.

Running notes on things I’ve hit, fixed, or found worth remembering.

My house is full of automation that never told me anything — until I gave it one push bus. The first thing I taught it to do was warn me before Claude Code cuts out mid-task.

I set out to answer a simple worry — is someone trying to get into my server? — and found the scarier question underneath it: if they did, would I even know? My front door was solid. The inside had an alarm with the wires cut, a web terminal sitting on the open internet, and no floor under the blast radius. Here’s the audit, and the three things I fixed.

Parliamentary roll-call votes are public, machine-readable, and almost completely unread. I built a thing that scrapes them, distills each decision into one plain-language question, shows which party voted which way, and lets you register whether you agree — then puts your answer next to how parliament actually voted. The rule that keeps it honest: the AI writes the summary, but it never decides a fact.

An AI agent on a scheduled idle walk through my notes pointed out that I’d built the same architecture three times — at work, in my homelab, and in my second brain — and that the third copy was missing the part that makes GitOps work. It was right. So we shipped the missing piece the same day.