A photo book built like a software release: data, checks, render, preflight A photo book built like a software release: data, checks, render, preflight

A Book Is a Build Artifact: 1,109 Trip Items to a Hardcover

TL;DR After a week at a backcountry cabin, I turned about 1,109 photos, videos and clips into a 94-page hardcover for my trip buddy, in about six days, as a build pipeline instead of a layout session. The stages: a media catalog, local speech-to-text, vision records for every item, sharp-frame harvesting from video, 1,229 fact claims run through verifier and skeptic agents, an equal-billing balance check, WeasyPrint to render, a custom pikepdf finisher to hit PDF/X-3, and a 28-check preflight. Final preflight: 0 failures, 13 warnings, each warning with a written reason. The interesting bugs were all data bugs: a camera clock 54 minutes fast, Whisper hallucinating text over wind, iCloud photos that were not actually on disk, and an AI that badged a fish off a receipt. If you can express a book as data plus a build, then review, regression tests and reproducibility come for free. That is the whole idea. The brief The trip is covered in another post, which was about the live page that ran during it. This one is about what came after. I wanted to give my buddy something physical, and I wanted it to feel like a bound volume of an old geographic magazine, not a photo-book template. ...

November 6, 2026 · 10 min · zolty
A handheld lake video turned into a looping macOS aerial screensaver A handheld lake video turned into a looping macOS aerial screensaver

Turning a handheld lake video into a macOS aerial screensaver

TL;DR I wanted a canoe-on-a-lake clip from my September backcountry trip as an Apple-style moving wallpaper and screensaver on the Mac Studio. macOS 27 has no “use my video” button, so Claude Code built one. The clip was 75 seconds of handheld 4K HLG at 120 fps. It was tone-mapped to SDR with VideoToolbox, locked to the far treeline with OpenCV ECC (correlation ≥ 0.991 on every frame) and cut into a 45-second loop with no visible seam. The tripod lock costs a 7.75% crop. The seam needed two fixes nobody would think of up front: a morph crossfade for foreground parallax, and -g 600 because ffmpeg’s default keyframe interval made the water pulse ten times a second. It’s installed as a real aerial by adding an entry to the per-user manifest Apple ships, with a LaunchAgent that re-adds it if Apple refreshes the catalog. The desktop ended up as the 48 MP still taken 14 seconds earlier, and the video runs as the screensaver. ...

November 2, 2026 · 14 min · zolty
A stack of security advisories with exploited flaws highlighted A stack of security advisories with exploited flaws highlighted

Fall 2026 security roundup: record patch counts and AI-found bugs

TL;DR The story of the last three months is volume. Microsoft’s October Patch Tuesday fixed a record 1,411 flaws, and that’s not counting the 348 fixed earlier in the month. The worst exploited bugs were the usual suspects: edge devices. Citrix NetScaler had two zero-days exploited for weeks before the September 27 patch. Browsers and phones got exploited bugs too: Chrome CVE-2026-87491 (the seventh Chrome zero-day this year) and Apple CVE-2026-86950. A Shai-Hulud-style npm worm got valid provenance signatures by releasing through the victim maintainer’s own CI, and parked persistence in .vscode and .claude directories. AI is on both sides now. It’s finding bugs faster than maintainers can patch them, and AI coding agents are themselves a target. Why I keep a list like this I run a homelab with a public blog, a k3s cluster, a pile of self-hosted apps and a coding agent with a shell. Every category below has a version that applies to me. This roundup is the list I worked through for my own stuff, cleaned up. It’s dated: everything here is as of late October 2026. ...

October 31, 2026 · 7 min · zolty
A self-hosted share host acting as a data-sharing layer A self-hosted share host acting as a data-sharing layer

My Share Host Became a Data Layer: ACLs, Forms, Comments, Status Pages

TL;DR I built a self-hosted share host in early September because a Claude artifact link was private to me and I needed to hand a page to specific people who have no account on anything I run. Access control was the first feature, not an afterthought. Authentication and authorization are split on purpose. oauth2-proxy with Google proves who you are and authorizes nothing. A small FastAPI app reads a per-folder ACL object from the bucket and decides what you can open. Grants change in about 15 seconds with no deploy. Pages that take input don’t post to a database. Every checkbox and typed cell becomes a typed operation that lands as a merge commit on a JSON spec in GitLab, authored by the person who made it. Git is the config layer. Every page has a comment rail. Comments have changed real work: a gear list, and the status page for my blue-team security harness. Long-running jobs get a status page that rebuilds itself from the job’s own data. The security eval ran on one across three days. A restore job I was babysitting republished every three minutes from a systemd timer. Access control came first The problem was boring and specific. I make a lot of HTML with Claude: plans, reports, checklists. When I built this, an artifact link was mine alone. There was no way to say “this person can open it and nobody else.” Emailing HTML files around is how you end up with four versions of the same checklist. ...

October 29, 2026 · 9 min · zolty
A single overloaded worker node among healthy ones A single overloaded worker node among healthy ones

One k3s node, four months of CPU starvation, three false fixes

TL;DR One worker in my home k3s cluster, k3s-agent-4, has been starved for CPU on and off since 2026-06-13. I’ve written about it in my cluster knowledge base more than a dozen times. It was declared fixed at least once and the declaration was falsified twice. “Quiet for a few weeks” turned out to be temporary. What the node showed: requests at 87 percent, limits at 629 percent of CPU, probes timing out, and Longhorn’s engine-image pod restarting hundreds of times. The alert I’d written for this could not fire on a node starved enough to silence its own exporter. Replacing it was the most useful change of the whole saga. The latest and most promising finding is physical: the Proxmox host under that VM clamps its CPU to 800 MHz under load (BD PROCHOT). As of early October that is open, not fixed. This is a post about the shape of a chronic problem, and how many plausible explanations can be true at once. The cluster in one paragraph The home cluster is four Proxmox hosts running k3s VMs: a control plane and four workers, with Longhorn for storage. (It used to have three control planes, until I collapsed it to one on 2026-09-05.) Workers have 6 to 10 vCPU each. The worker in this story lives on the fourth host, the only one with hyperthreading: an i7 with six physical cores and twelve threads, hosting the 10-vCPU agent. ...

October 26, 2026 · 10 min · zolty
A set of classifiers each routing decisions toward allow, deny or a human A set of classifiers each routing decisions toward allow, deny or a human

A field guide to the classifiers in my stack

TL;DR I count six classifiers in my daily stack. Only three of them involve a model. The other three are keyword lists and parsers, and they are the ones I trust more. The useful question for any classifier is not “how accurate is it” but “which way does it fail?” Every one of mine fails toward a human, or toward the cautious side, on purpose. Four principles I keep relearning: deterministic checks first, fail toward a human, a classifier may only tighten, and measure against a golden set instead of vibes. The most common classifier bug in my setup was classifying text instead of the thing the text describes. Measurement is uneven. I have real numbers for some of these and none for others, and I have tried to say which is which. Why a field guide I did not set out to build a classifier collection. They accumulated, one incident at a time, each time something needed to decide “is this one of those” before an action happened. Resale is the hobby that funds the hobby, and a lot of the automation around it is the same question asked in different places: should this thing be allowed to proceed? ...

October 22, 2026 · 10 min · zolty
A small golden set of lots scored with paired comparisons and confidence intervals A small golden set of lots scored with paired comparisons and confidence intervals

Benching a vision pipeline on 17 lots without fooling myself

TL;DR My pipeline reads photos of surplus lots, finds comparable sales, and does arithmetic to produce a bid ceiling. To change any part of it safely, I need a measurement that can tell a real improvement from luck. The method: a small golden set of lots with human-verified answers, paired comparisons (every candidate sees the same lots), an exact McNemar test on the cases where candidates disagree, and Wilson intervals on every rate. On 17 lots, “accuracy %” is noise. A single lot moves the number by about six points, and the 95 percent interval on a respectable-looking score is roughly thirty points wide. Reading make and model is the easy part. Counting units and choosing comps is where the error enters, so I score those separately from identification. The measurement that actually changed my mind was not a benchmark at all. It was re-running a batch of lab lots by hand: the summed bid ceilings dropped by about 72 percent. No per-model leaderboard here, on purpose. The method is the transferable part, and a small bench makes model rankings mostly an artifact. This is the method behind the numbers in the identification post. Where that post shows how the pipeline works, this one is about how I know whether a change made it better. ...

October 19, 2026 · 9 min · zolty
Reading option logits from a small local model Reading option logits from a small local model

Jev in 25 lines: building the decision model locally

TL;DR In the design review I concluded that Jev, TypeSafe AI’s typed decision model, is cheaper and faster than an LLM and not more accurate. That raised an obvious build-vs-buy question: if the value is “cheap probability over a fixed set of options,” can I get that from a small open model running on hardware I already own? Mostly yes, with one catch that matters. The technique is to prompt with lettered options, run one forward pass, and read the logits for just the option letters. A parody post called “Jev in 25 lines of Python” does exactly this, and the technique under the joke is real. The catches are calibration (raw softmax is bad until you fit it) and serving (Ollama could not do it when I wrote the design). ...

October 16, 2026 · 8 min · zolty
A gallery of surplus lot photos feeding a structured identification schema A gallery of surplus lot photos feeding a structured identification schema

Reading a surplus lot from its photos: the identification pipeline

TL;DR Resale is the hobby that funds the hobby, and the expensive mistake in it is not a bad price. It is a confident wrong model number, because everything downstream (comps, margin, bid ceiling) is built on it. Auction search results give you one thumbnail and a title the seller typed in a hurry. The real information is in the gallery: wide shots for counting, close-ups for labels. The pipeline is split on purpose: vision reads the lot, a comps step prices what was read, and plain deterministic arithmetic turns that into a number. The model never does the math. Fetching the full gallery moved make/model accuracy from 0.375 to 0.875 on a 24-lot hand-labelled bench. That is the single biggest lever I have found, bigger than any model swap. Most of what went wrong later was not misreading a badge. It was counting and itemizing: accessories sold as products, “X with Y” counted twice, a screen size parsed as a quantity. A final sanitizer can only tighten a verdict. And if the gallery fetch fails, the lot is deferred, not guessed. I wrote about how lots get found in the first place. This post is about the step after: given a lot that looks interesting, how do I know what is actually in the pile? ...

October 13, 2026 · 11 min · zolty
Security model evaluation results Security model evaluation results

Grading Seven AI Models on Blue-Team Security Work in a VM With No Network Card

TL;DR I wanted to know which models I can trust for the defensive side of running a homelab: reviewing code for vulnerabilities, reviewing Kubernetes and Terraform config, triaging logs, and reading threat intel without making things up. So I built a harness with 112 frozen synthetic cases, graded everything deterministically (no LLM judge), and ran every model-written fix inside a Proxmox VM whose network card was deleted at the hypervisor. ...

October 12, 2026 · 16 min · zolty

Affiliate Disclosure: Some links on this site are affiliate links (Amazon Associates, DigitalOcean referral). As an Amazon Associate, I earn from qualifying purchases. This does not affect the price you pay or my editorial independence — I only recommend products and services I personally use and trust.