How a repair-kb skill and a bench-troubleshoot skill let me work on an HP 3458A, HPLC modules, vacuum pumps and balances I had never opened, and where the AI was wrong.
Badges
Badges come from GitHub once the author claims this feed.
Recent posts
every link opens the author’s post-
I fix lab gear I've never seen by having Claude build the repair knowledge base first
-
The office screensaver is a news wall, and it broke in six ways
An idle-watch screensaver on office all-in-ones that rotates headlines, markets, radar and sky.
-
There is no best personal AI assistant. I built one, then retired it.
OpenClaw, Claude, ChatGPT Work, Gemini Spark, Siri, Hermes and more, with pros and cons.
-
A Book Is a Build Artifact: 1,109 Trip Items to a Hardcover
How a pile of trip photos, videos and voice notes became a 94-page print-ready hardcover in about six days, with a fact-check pipeline and a PDF/X-3 finisher.
-
Turning a handheld lake video into a macOS aerial screensaver
How a shaky 4K iPhone clip became a native macOS 27 aerial: HLG tone mapping, ECC stabilization, a ghost-free loop seam and a hand-built aerial manifest.
-
Fall 2026 security roundup: record patch counts and AI-found bugs
NetScaler zero-days, exploited Chrome and Apple bugs, a provenance-signed npm worm, AI coding tools as targets, and a record 1,411-fix Patch Tuesday.
-
My Share Host Became a Data Layer: ACLs, Forms, Comments, Status Pages
A self-hosted share host I built for access control grew forms that commit to Git, comment threads that steer work, and status pages for multi-day jobs.
-
One k3s node, four months of CPU starvation, three false fixes
A single worker kept throttling Longhorn, MetalLB and node-exporter.
-
A field guide to the classifiers in my stack
Six classifiers I run or live under: a permission gate, a command guard, a lot filter, a reply sorter, a goal judge and an MR reviewer.
-
Benching a vision pipeline on 17 lots without fooling myself
How I measure a lot-reading pipeline with a small golden set, paired McNemar tests and Wilson intervals, and why an accuracy percent on 17 lots is mostly noise.
-
Jev in 25 lines: building the decision model locally
Reading option-letter logits from a small local model gives you a Jev-shaped classifier for free.