This Site: Designed & Shipped with AI

I chose Astro over a drag-and-drop builder on purpose. Astro ships static HTML with zero client-side JavaScript by default, which is the honest way to make a content site fast. It also gave AI a real, componentized codebase to work in instead of a locked template. TypeScript everywhere, deployed on Vercel. The result is a site I can reason about, version in git, and hand to any developer. Same standard I demand of a node graph.

The typography is self-hosted. Poppins and Lato are served from the project itself rather than fetched from a font CDN, so the look is locked, private, and fast. Fluid clamp() sizing means every headline scales cleanly from phone to desktop without a rack of breakpoints. AI drafted a lot of this CSS; I set the scale, the rhythm, and the constraints it had to stay inside.

The site isn't a stack of one-off pages. It's a small design system. Every case study is composed from the same reusable Astro components: SectionLabel, Figure, Callouts, Stats, PullQuote. This page is built from the exact set the others use. Content lives in single-source data files (work.ts, site.ts, covers.ts), so a fact like my contact email or the order of the work grid is defined once and read everywhere.

That structure is a deliberate directing decision. AI is at its best filling in a clear pattern and at its worst inventing architecture on the fly. So I set the pattern (components in, data separated out, one source of truth per fact) and let AI move fast inside it. When it drifted, a hardcoded string that belonged in a data file, a bespoke layout that should have reused a component, I pulled it back. The node graph is the new creative brief; here, the component tree is.

"Fully deployed" means more than pushing to a host. This site runs on real production infrastructure, and I built the operational layer myself, with AI as the pair and my judgment on every security-relevant call.

This is the section prompt-and-pray portfolios never write. AI got real things wrong, and catching them is exactly the accountability I teach.

The disappearing spaces. An AI-suggested config turned on HTML compression, which quietly stripped the newline between text and inline tags like <strong> and <a>, deleting the space between words wherever a line wrapped in the source. Copy read fine in the editor and broke on the page. The fix was deliberate: compressHTML: false, with a comment explaining why, so no future edit silently re-breaks it.

The security model. The first pass at the preview gate would have signed session cookies with the site password itself, which turns a captured cookie into an offline password-cracking oracle. I directed the rewrite: an independent signing secret, fail-closed behavior when it's absent, and a deliberate delay on failed attempts. AI wrote the code; the threat model was mine.

AI-generated portfolio sites are everywhere now. Some of them look stunning: incredible animations, smooth page scroll effects, interactions that wow on first load. Then you look deeper. Speed? Neglected. Accessibility? Skipped. Security? Never considered. Fact-checking? Zero. The codebase underneath? Bloated and heavy. It's prompt and pray at its finest: someone who "knows AI" tells the software to "make a cool website with lots of motion and animations," watches the first output, and ships it without ever reading what the AI actually produced. The site looks like the designer knows their craft. The code says otherwise.

Those sites treat AI as a finish button, not a tool inside a production pipeline. The person using it never asks: Who is going to use this? How are they going to use it? Where is this product going? Is the codebase I'm shipping light, medium, or unnecessarily heavy? What happens when a mobile user on 3G hits this page? Can someone navigate this without a mouse? If I need to change a fact or fix a bug, can I do it without calling the AI back in? They skip all of that. They get animations and call it done.

The site reports its own vitals from production via Vercel Speed Insights. Every number below is field data from real visits, not a lab run, and the page weight is measured transferred bytes. No score I can't show you the source of.