I've always been drawn to what's technically possible right now, at the edges - before it becomes obvious. So when a real operational problem crossed my path, I didn't reach for an off-the-shelf solution.
As co-founder of the Usability Testessen Berlin chapter, I'd been running a meetup format that's genuinely unusual: a structured speed-dating setup where participants are matched against products to test in several rounds, not just randomly seated. That means the logistics are more complex than a typical event. Invitations go out, applications come in, attendees need to be selected to hit a precise ratio of testers to products, no-shows have to be accounted for, running orders generated, and on the night itself, check-ins managed in real time - including decisions about whether to promote backup visitors to active testers and strikes for no-shows. For years, we handled this manually. It worked, but it was slow and error-prone.
So I built a tool to fix it.
How I built it
Before writing a single line of code, I did architecture sparring with an LLM - not to let it design the system, but to pressure-test my own thinking. I wanted something modular, maintainable, and fast. Not AI slop.
From there, I vibe-coded a React frontend and backend, paired with a PHP-based API. The API choice was deliberate: most LLMs aren't optimized for PHP, which meant I had to stay in the driver's seat. That was the point. I wanted code I could actually read and own.
The development process itself was a lesson. Vibe coding is genuinely different from classical product design: the cost of trying something is so low that iteration becomes the default mode. That's powerful - but without clear direction, it produces feature sprawl fast. I had to keep steering actively.
Once the core was working, I shifted focus to hardening. The tool handles personal data for hundreds of people and is publicly accessible, so security wasn't optional. I ran back-and-forth pen testing sessions with LLMs alongside my own review process until I was confident.
Then: real-world testing across multiple events, with continuous feedback from the whole team.
Where it stands
- Planning time is down roughly 80%.
- After just three events with the tool, we've processed over 1,000 applications - without a single major error.
- Development is ongoing and the tool gets better with every event.
The less obvious result: working through this process gave me a much sharper read on where vibe coding actually delivers and where it breaks down. An idea of what vibe coding can archieve and where it might be overhyped. That's not just a side project insight - it directly shapes how I guide my design team's work with AI tools today.
