Blog — Arclight

Sep 292026

Pulsar 26.9.1: One Theme for the Whole Desktop

Pulsar, my nightly-built Fedora Silverblue desktop, gets its first numbered release: a theme engine that recolors the shell, apps, terminal and wallpaper in one step, translucent shell menus, and 20 themes.

Disposable by Default

Agents made building cheap and maintenance exactly as expensive as before. Most of what I build now gets deleted. Here's the bar for what gets to stay.

The Camera Is Part of the Model

Most production CV accuracy problems are capture problems. Why the physical setup deserves engineering time before the architecture does.

What a 4,400-Test Suite Actually Buys You

ARC UI 3.0 replaced manual review with strict automated gates. What that cost, what it caught, and where it still lets things through.

Generating Framework Wrappers Instead of Writing Them

One Lit source, six framework bindings. Why Prism generates ARC UI's React, Vue, Svelte, Angular, Solid, and Preact packages rather than maintaining them by hand.

Every Tailwind Take You've Read Is Wrong

Class soup isn't a problem. Config lock-in isn't a problem anymore. There's one thing worth watching, and it's on you, not the tool.

Detection vs. Segmentation: When the Pixel Mask Is Worth It

Bounding boxes are cheap and usually enough. A client-facing guide to when segmentation earns its extra cost in inspection pipelines.

The Eval Set Is the Product

Models get replaced. The labeled test set that judges them outlives every one of them, which is why it deserves more care than the training run.

Classical ML Still Ships

SVMs and Mask R-CNN-era models keep winning on latency and total cost of ownership in production inspection systems. When to reach for them over the newest architecture.

Ship Code, Not Roadmaps

Milestone pricing, a staging environment from day one, and a hard cap of 20 hours a week. Why Arclight's constraints are the deliverable.

When Not to Use AI

A pragmatic guide to recognizing when simpler tools outperform machine learning, and why knowing the difference matters.

Ethics in AI: The Right Way Is the Hard Way

Why I avoid diffusion models, train exclusively on authorized data, and believe AI works best when applied surgically, not as a blunt instrument.

Announcing ARC UI: 166 Web Components, Open Source

I built my own web component library from scratch. ARC UI is Lit-based, ships native wrappers for seven frameworks via my Prism code generator, and is now open source at arcui.dev.

Less Is More: Why I Reach for Simpler ML Models First

Why SVMs, Mask R-CNN, and classical ML models still outperform deep learning in production, and how to choose the right model for accuracy, latency, and total cost of ownership.