I lead an AI engineering team in Berlin. We build agents that people rely on at work. In my research, I look at what you can verify when all an agent gives you is its account of what happened.
AI Reliability Layer
I built a shared system that spots failures across our production agents and recovers failed runs.
Agent Cheat Index
Did the agent fix the code or change the test? Compare 10 coding-agent setups across 1,162 runs.
Receipts, Not Transcripts
In a team study, we compared agents’ logs with the server record. Their logs left out 573 of 8,211 operations.
Code review: vLLM Semantic Router. I flagged duplicate score IDs; the author added validation and regression tests.




