How this site is built, and who builds it

This site is a project like the others: it has a stack, opinions, a test suite, and a few scars. This note is the colophon, plus an honest answer to a question people have started asking: how much of this did AI build?

the stack

Next.js 16 with the App Router, TypeScript throughout, Tailwind 4 with a CSS-first config. Type is Fraunces for headings, Geist for body, JetBrains Mono for data and labels. The globe on the travel page is React Three Fiber. Hosting is Vercel.

The live parts pull from real sources through 15 API routes: Spotify for what I am playing, GitHub for what I am shipping, Apple Health data posted by an iOS Shortcut into Vercel KV, and a training payload published by my marathon pipeline. Everything else is flat files in the repo, JSON and MDX, because for a site maintained by one person, content in git beats a database.

the design rules

Content first. Data and interaction only where they add something. Lowercase labels, generous whitespace, one accent color, dark mode that gets the same attention as light. If a page makes someone work to understand who I am or what I made, the page is wrong.

the tests guard the truth

The part I actually recommend copying: the test suite protects facts, not just code.

Unit tests cover the data and parsing layers, 61 at last count. Content integrity tests fail the build if a screenshot goes missing, an internal route breaks, or a project's proof points drift from what the repos can back up. Playwright drives the real built site and runs axe accessibility checks against WCAG A and AA on every core page. A live-check script hits production after deploys and fails loudly if a demo is down or if copy that should not exist creeps back in. Security headers are set explicitly, admin routes sit behind cookie auth, and the fitness data goes through a field whitelist so GPS and raw health detail never deploy.

The résumé here also exists for machines: structured data for the pages, a generated social image per project, and an llms.txt index for whatever reads sites now.

who builds it

I design and decide. AI agents do an increasing share of the typing, and I think being specific about that arrangement is more interesting than hiding it.

The repo carries instructions for agents, including a warning that this Next.js version is newer than their training data and that they must read the bundled docs before writing code. The test suite is the leash. An agent can propose anything, but the accessibility checks, the content integrity tests, and the truthful strings asserted in end-to-end tests decide what ships. When an agent rewrote parts of this site, the tests are what kept every number and claim honest.

That arrangement, judgment stays with me, verification lives in code, labor is shared, is the same one my marathon coach runs on. It seems to be the shape I keep arriving at. I wrote about the deeper version of it in my marathon coach is a chat window.

still changing

The site keeps growing sideways: the fitness page turned into a live dashboard, travel turned into a globe, and now writing has its own corner. That is the fun of it. It can keep changing as long as the tests keep passing.