DE
Esc

← All projects

Looking back, written the day after the hackathon.

Voya: 48 hours, more than 50 worktrees, no prize

At the OpenAI hackathon at IdeaLab! in Vallendar I coordinated parallel AI coding agents. We didn't win, but the MVP is live.

No prize. The MVP is live anyway, and I worked differently in those 48 hours than ever before.

The setting: an OpenAI hackathon at IdeaLab!, the startup conference of WHU in Vallendar. René and a designer were on my team. Our idea came straight from the conference itself. Founders there are surrounded by hundreds of profiles and have no idea which conversations are worth having. Co-founders, investors, mentors.

Voya turns that into a flow. An agent interviews you, by text or by voice through OpenAI’s realtime API, and builds a search profile from it. Then it ranks possible contacts with a score that explains why someone fits instead of just spitting out a number. For reaching out there are message drafts adapted to the other person’s personality type, and a voice agent you can role-play the conversation with beforehand. Everything also runs without an API key, on rule-based fallbacks, so the demo doesn’t die on a network issue.

The demo on this page replays exactly that flow, with invented sample profiles instead of real conference data: the interview, the search profile, the ranking with a score you can expand, and a message draft per personality type. The score is deterministic, so the same inputs always give the same result, and it comes with a reason.

voyaCo-Founder Studio
Dein privater WorkspaceD
Dein nächstes Kapitel Das passende Team

Zusammen mehr bewegen.

Entdecke die Menschen hinter den Lebensläufen.

Aus deinem IdeaLab-Netzwerk
8 Profile · nach Passung sortiertRegelbasiert berechnet, ohne LLM
01
02
03
04
05
06
07
08
Demo with sample data
Studio shell after René’s Voice Studio; matching, match breakdown, outreach and role-play from our Next.js app. All rule-based, no LLM; every person is invented.
Fig. Voya, from the interview to the role-play Plays by itself · click to take over

What was new for me was the way of working. I took over coordination and cut the work into more than 30 packages, each with clear ownership of files. Then AI coding agents worked on them in parallel in more than 50 git worktrees. My job was less typing and more slicing, checking and merging. Things I wrote myself include the match API with explanations, the outreach templates, the agent core, the voice integration and the design system. In the end 130 of 142 commits were mine, live on Cloudflare Workers.

The rule I took from it: two agents must never touch the same file. That’s what the ownership packages are for.

Tools and skills: Next.js 16 with TypeScript and Tailwind, the OpenAI Agents SDK, the Realtime API over WebRTC for voice, zod, Cloudflare Workers, our own design system with tokens, git worktrees, and cutting work into packages that agents can build in parallel.

We didn’t win. Since yesterday I’m full-time at Kobe, and the topic isn’t far from Voya: finding the right people.