How this site was built
This site is also the proof. It was designed and built solo, start to finish, with an AI coding assistant in the loop the whole way — which is how Paul builds everything now. Here is the honest version of what it is made of and the calls that shaped it.
The stack
Lean on purpose. No framework soup, no backend to babysit.
- Next.js 16 (App Router)
- Static case studies and writing, plus route handlers for the live AI.
- React 19 + TypeScript
- Typed end to end — every component and every API contract.
- Tailwind CSS v4
- CSS-first theming, no config file. One token set, used everywhere.
- Anthropic SDK — Claude
- Claude Sonnet runs the ask-bar and the role matcher. Streaming over web standards.
- Content pipeline
- gray-matter, react-markdown, remark-gfm and rehype-raw render the posts and case studies.
- next/font + next/og
- Self-hosted Saira, Space Grotesk and JetBrains Mono, plus a generated share image.
- Vercel
- Hosting and the serverless functions behind the AI. No separate backend to run.
- Vercel Blob (private)
- A private store holding the anonymized question log — readable only with a server-side secret, never public.
Built with AI, honestly
The design happened in real, rendered mockups instead of static comps. The build happened component by component, with Paul making every design and architecture call and an AI assistant doing the heavy lifting on implementation.
The split is simple. Paul owns the judgment, the taste, and the decisions. The AI accelerates the parts that used to need a team. The back end he never used to touch got handled. The front end he always built got faster. He doesn't write every line. He steers it, tunes the details, and catches where the model drifts. Nothing ships until it meets his bar.
The calls that mattered
- 01
One model, on purpose
The ask-bar and the role matcher both run on Claude Sonnet. Both speak for me when I'm not in the room, so they get the sharper model. Accuracy matters more here than saving a few cents per answer.
- 02
The matcher is built to be honest
It returns structured output through forced tool use, validated against the real project list, and it is tuned to flag genuine gaps instead of overselling. A tool that lies about fit is worse than no tool.
- 03
The assistant talks about Paul, never as him
Third person on purpose. An AI pretending to be the person reads fake. A guide that presents evidence and lets you decide reads honest — and it shows the engineering instead of hiding it.
- 04
The work never waits on the model
Tap a question and the work canvas re-curates instantly, client-side, while the written answer streams in behind it. If the AI is slow or capped, the visual response still lands.
- 05
A public AI endpoint needs a ceiling
Per-IP rate limits plus a hard monthly spend cap. The site can invite anyone to use the AI without anyone being able to run up the bill.
- 06
Log the questions, never the people
The assistant keeps a private, anonymized log of what gets asked — the question, a coarse city, and a one-way hash of the IP so repeat visitors can be told apart. The raw IP is never written down. Enough signal to learn what people actually want to know, without tracking anyone.
- 07
The words are human
The copy was written and rewritten to sound like a person, not a model. This site leans on AI heavily under the hood and almost not at all in its own voice. That was the point.
Open the hood
None of this is a black box. This page is the tour of exactly how it's built, the assistant answers stack questions straight, and the rest of Paul's work is already out in the open.
github.com/somekidpaul