Cooking foundational AI and robots for food at KAIKAKU. Born and raised in Austria, currently in London.

Peeled potatoes at my parents’ restaurant since age 6. Then built one of the world’s first bitcoin faucets at age 13 with 150k daily users, which somehow led to leading software in the Austrian military. Eventually the food won.

Research

Papers on food embeddings, physical AI, and when combining language models actually helps. The Epicure work trained ingredient embeddings on 4.14 million recipes; the sketch below is a hand-placed miniature of that map. Drag it around. The dashed arcs are pairings the model swears by, like caramel and fish sauce.

• When Does Combining Language Models Help? A Co-Failure Ceiling on Routing, Voting, and Mixture-of-Agents Across 67 Frontier ModelsJun 2026

  Combining LLMs rarely beats the single best model. A co-failure ceiling, measured across 67 frontier models.

• AEGIS: A Backup Reflex for Physical AIJun 2026

  Detecting imminent manipulation failure from a weak policy's internal states, then switching to a stronger one just in time.

• Epicure: Navigating the Emergent Geometry of Food Ingredient EmbeddingsMay 2026

  Ingredient embeddings trained on 4.14M recipes, from chemistry to recipe context.

All 7 papers →

Writing

Braindumps. The latest three:

• Jul 7, 2026Turks All the Way Down
• May 29, 2026The Genghis Khan Operating System
• Feb 25, 2026Cathedrals and Europe's Case for Forward Deployed Engineering

All 8 braindumps →

Investing

$5k-100k angel checks into 50+ pre-seed rounds, mostly founders building physical things: Isembard, Dirac, Minerva, and friends. Usually the first or weirdest check on the cap table.

Full portfolio →

Taste