I gave my second brain two agents that read the outside world and collide it against my notes. A Magpie watches my GitHub stars and only speaks when something hits live work. A Blue Jay reads a handful of RSS feeds and surfaces the distant, not-yet-relevant connection. They share a security spine — and they have deliberately opposite jobs. Here’s why the split is the whole design.
My second brain searches over a vault of markdown using BM25 + vectors + graph expansion. I’d been telling people the graph improved recall. Then I finally benchmarked it, and plain keyword search beat my fancy hybrid — the graph was actively dragging the right answers out of the results. Here’s the scorecard and what it taught me about where graphs actually belong.
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.
I built a clock that counts down to a form of government that doesn’t exist yet — legitimacy grounded in verified knowledge rather than power, wealth, or whoever shouts loudest. The only reason I’m not embarrassed to have built it: the clock can run backward, the assumption behind it is published in plain sight, and the first concrete brick already ships real parliamentary data. A measurement, not a prophecy.
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.
I asked Claude Fable 5 to help me self-host a dashboard for my own Claude usage. Halfway through, its dual-use safety measures flagged the conversation and downshifted me to Opus 4.8. Nothing I did was wrong — the request just had the shape of something that is. That gap, between what a thing looks like and what it’s for, turns out to be the whole story.
Zero-shot voice cloning with XTTS-v2 on a CPU-only k3s node: 26 seconds of phone audio in, a cloned-voice audiobook out — and an honest verdict from the bedtime jury. Every manual step, including the ones that went wrong.
I gave my markdown knowledge base a nightly gardener — an AI that finds orphan notes and missing links and fixes them, every change a reviewable git commit. The fun part was the Kubernetes wall I hit on the way.
You don’t have to be job-hunting to want to know your market — what’s out there, what it pays, where you’d fit. So I built an n8n workflow: it polls the public ATS APIs (Greenhouse/Lever/Ashby) plus a broad remote-jobs feed, filters for remote-EU infra roles, scores each posting against my CV with an LLM, and emails me only the 80%+ matches. No database, no scraping.
My first second brain died the way most do — on multi-device sync. The rebuild: plain markdown as the source of truth, every clever layer derived and disposable, and an AI that tends it through reviewable git diffs.