What I do
I pick problems where AI can do work at a scale people can’t, and then I make that work trustworthy. On InfiniteGrammar.de that meant choosing a narrow problem (deep grammar practice on one topic at a time), defining what a good exercise is, building the system that generates and checks them, and watching what learners actually do with the result.
I work with AI agents as my engineering team. Claude Code writes most of the code from specifications I write and review. I keep the decisions, the budgets and anything that touches production data. That way of working is described in detail in this article.
Selected InfiniteGrammar.de milestones
| Milestones | Articles |
|---|---|
| Drafts passing every quality check: 17% → 78% on held-out tests | Case study |
| Cost per accepted exercise: $0.23 → $0.058. Cost per clean exercise $0.089, against $0.372 for the configuration before | Unit economics |
| Accepted exercises graded flawed in blind audits: 28% for an early version → 0% (0 of 31) today | Decisions |
| Whole live library re-checked and replaced: 974 new exercises in four days for $78 | Live audit |
| Organic search clicks +77% after the first wave of agent-made SEO fixes | SEO agent |
Background
I’ve spent over a decade building digital products in Berlin, with a particular focus on data, machine learning and AI. Along the way, I’ve built ML-based pricing, fraud detection and segmentation tools, worked hands-on with ML models and LLMs, and led complex platform projects. I started out as a software developer, studied computer science and business administration, and spent five years teaching Scrum and product management.
Contact
If you’d like to see something in more depth, such as the decision log, the test sets or the data contract behind the dashboards, ask. I’m happy to walk through any of it.
The quickest way to reach me: