Skip to content

Under the hood

How this site is built.

A portfolio should prove skills, not just list them. Here is the architecture, the guardrails of the AI assistant and the checks that run on every change.

(01)AI

The AI assistant, step by step

  1. (01)

    Instant local search

    Every project, experience and school is indexed in the browser. Typing in the ⌘K palette ranks results immediately, with no network call and no AI.

  2. (02)

    Grounded generation

    For a real question, the server sends the whole portfolio as context to a Mistral model with a low temperature (0.2). The prompt forbids inventing numbers, companies or dates.

  3. (03)

    Guardrails

    Questions are capped at 400 characters, history at 6 messages, answers at 400 tokens. Each IP is limited to 15 requests per 10 minutes and the upstream call times out after 25 seconds. Instructions hidden in a question are ignored.

  4. (04)

    Graceful degradation

    If the AI key is missing, the quota is reached or the provider fails, the assistant falls back to the local search results instead of showing an error.

(02)Stack

Architecture choices

  • Static first

    Next.js 16, React 19 and Tailwind CSS 4. Pages are pre-rendered at build time; only the assistant is dynamic.

  • Content as data

    Projects, experiences and profile are typed YAML files edited through Keystatic. Every edit is a Git commit, so the content has a full history.

  • Bilingual by design

    English at the root, French under /fr with translated URLs, hreflang alternates and a sitemap for both languages.

  • Readable by machines too

    llms.txt and llms-full.txt expose the portfolio to AI crawlers; structured data (JSON-LD) describes each page for search engines.

(03)CI

Checks on every change

  • ESLint with the Next.js rules
  • Strict TypeScript type-check
  • Full production build (all pages in both languages)
  • Accessibility basics: skip link, focus states, reduced-motion support, labelled controls