We didn’t win anything. Let’s get that out of the way.

On March 17 at 4 am, René and I got back to Berlin from the road trip. The same day I left for St. Gallen. With Ignite, the student founders’ initiative at HPI, twelve of us went to START Hack in three teams, with travel funding from HPI. START Hack calls itself Europe’s most entrepreneurial hackathon: more than 600 people, 36 hours, challenges from partner companies.

Our team was called Thesinator. The challenge came from Studyond, a Swiss startup: design and prototype an AI-guided journey through your thesis. With one clear rule. The AI suggests, the professor decides.

We built an entry flow with three questions, voice included, topic matching with embeddings on Supabase and pgvector, a multi-agent simulation that shows how a topic might play out, and a chat with your future self. My part was the chat, networking and browse features: more than 20 React components, two Supabase Edge Functions with their database migrations, and the persistence that lets the entry flow survive a reload.

One thing was new for me. Before we started, I condensed the challenge’s 89 context files into a compact knowledge base for our coding agents, one file they read first. The idea: the agents shouldn’t need the problem re-explained in every prompt.

The demo on this page shows how it fits together, with sample data. In the wizard you answer three questions, typed or spoken. Your answers become an embedding, a vector of numbers that describes what you’re after. pgvector finds the topics in the database whose vectors are closest, and they show up as cards you swipe left or right like in a dating app. After that the network view opens, with people, organisations and topics you can browse and ask questions about in a chat. Those network, browse and chat parts are what I built. Two Supabase Edge Functions answer the chat, and the session is saved, even if you reload in between.