🔥 Founder @ CodeBurn · 📍 Germany
7 years of writing embedded C++ for self-driving cars. A detour through Web3. Now writing probes that break AI agents before someone else does. Open source by default, because most AI security tooling costs more than the agent it's protecting.
CodeBurn tracks token usage, cost, and performance across 40 AI coding tools.
Durable, local-first memory for AI agents. Not a vector store: Eywa distils conversations into atomic facts, each carrying the verbatim quote it came from, links them into a knowledge graph, and supersedes stale values as the world changes. It consolidates in the background while nothing is running, like dreaming. High-stakes facts keep their provenance, so a poisoned memory gets caught instead of obeyed. Agents wake up knowing what they knew.
- 3 to 4 Claude Code sessions in parallel, every day. The panel below is what that looks like.
- A research team of AI agents: they surface new attack surfaces, file the tickets, work the research, open the PRs. I review, they iterate. Nothing merges without my word.
- A separate marketing agent per product.
- 10+ hour days. Germany. Mac only. Coffee.
- Agentic engineering, not vibe coding. Seven years of debugging embedded C on autonomous vehicles taught me exactly where things break. AI lets me ship faster. It does not let me ship sloppy.
measured, not vibes: my own usage, tracked by CodeBurn, updated nightly
Keep trying.
Nothing I've shipped looked like a hit when I started it. CodeBurn was a five-minute reply to a Reddit complaint, built in two hours, pushed the same day. It's past 9,500 stars now. If I'd waited for "ready," it wouldn't exist.





