A portfolio for humans and machines
How I built this website – static, bilingual, with a strict CSP and a terminal that ships no data model to the browser.
A portfolio is no longer read only by people. Recruiters ask their AI assistant, search engines parse structured data, and agents call APIs directly. This site is meant to be unambiguous for all three. These are the decisions behind it.
One source of truth, checked at build time
My CV, skills and projects are written exactly once in the repository. Every page, every endpoint and the JSON résumé read from that one source. Zod validates it during the build:
export const localized = z.object({ de: text, en: text });If a text is missing in one language, the build fails. A mistake never reaches the website, only my CI.
Two languages, enforced
German is the default at /, English lives at /en. Every post has to exist in
both languages – including this one. The build checks that too, instead of relying
on discipline.
Static, with a strict CSP
The pages are plain HTML, built with Astro. A hash-based Content Security Policy allows only the scripts I ship myself. That is why I use the browser’s native view transitions instead of a client-side router: they need no JavaScript and work with the CSP.
A terminal with no data model in the browser
The terminal on the start page feels dynamic but knows no data. Every answer is computed during the build and embedded as JSON. What reaches the browser is a lookup table of about two kilobytes – no validation code, no data models, nothing that can drift apart.
What comes next
Next, the site gets an MCP endpoint. Agents will then be able to query my profile and propose an intro call without reading a single line of HTML.