This site: a portfolio with an AI agent
A portfolio built as a desktop: folders, windows, projects with screenshots and a live agent that answers questions about my experience from my own texts.
Screenshots
Projects section 
The agent answers from the site's own texts 
Project card 
Screenshot viewer 
Case study window 
The agent on a phone 
Desktop on a phone
Instead of a usual resume landing page, there is a desktop: folders with case studies, windows you can move around, a Projects section with screenshots and a chat with an agent at the entrance. The agent answers a visitor’s questions about my experience, looks for the answer in my own texts, and if it does not know, offers to leave a contact.
The frontend is Astro with Preact islands. Pages are static, all the JavaScript they load fits into 17 KB, and a CI test checks that. Every window also opens as its own page, so each case study has its own URL, og image, sitemap entry and a place in llms.txt.
The agent runs on Fastify and streams answers over SSE. It has a model spend cap, per-IP limits and prompt-injection protection. Context for each answer comes from a Python RAG service: hybrid search over embeddings and BM25 in SQLite. The case study below explains how it works.
Since July 2026: over 230 commits and about 800 tests across the site build, the agent server, RAG and the admin panel.