Colophon

Stack

Astro 6.3

Static framework + one SSR endpoint

Preact 10

Two interactive islands (chat, contact form)

Cloudflare Workers edge runtime

Host - prerendered HTML plus /api/chat

LangChain 1.x (core + openai)

Chat message plumbing

OpenAI gpt-5-mini

Chat model

Latin Modern Roman Computer Modern revival

Typeface - the LaTeX-preprint reference

Bun 1.3

Package manager and script runner

Decisions

1. 01

   Astro + Cloudflare Workers, not Next.js + Vercel

   Prerender everything except the chat endpoint. One SSR route. Static assets and the Worker share a single deploy. The site costs nothing at idle and serves HTML from the edge cache.

2. 02

   Preact, not React

   Two interactive components on this site. Preact ships ~3 kB of runtime; React would ship ~45 kB. The DX is identical - JSX, hooks, the rest. The choice is invisible to a user; it shows up as a faster first paint.

3. 03

   Two-layer prompt-injection defense

   The system prompt explicitly tells the model to treat <user-input> contents as data, not instructions. The wrapper escapes any literal closing tag a visitor might type, so they cannot forge a fresh role boundary. Neither layer alone is enough; both together hold.

4. 04

   Server-stateless chat

   The conversation transcript lives in the browser; the server holds no session state and caps history to the last five exchanges. Each request is logged to D1 (question, reply, salted IP hash) for rate limiting and post-hoc improvement, and purged after a year - the privacy page spells it out.

5. 05

   Per-page prerender by default

   `output: 'server'` with `export const prerender = true` on every page except /api/chat. Static where it can be, dynamic where it must be - the configuration mirrors the architecture, not the other way around.

6. 06

   Readable by machines on purpose

   AI crawlers are welcome here: robots.txt allows them by name, and /llms.txt plus /llms-full.txt serve the whole site as plain text in the llmstxt.org format, generated from the same data files as the HTML. If an AI answers a question about Ehsan, it should have the facts.