🏃 Runner · 🛠️ Building AI tools I actually use · 🧱 Architecture over shortcuts
I spend most of my free cycles on agent engineering — turning LLMs into things you'd trust with a real decision, not just a demo. CLI-first, composable, verifiable.
A safe, agent-friendly bridge between LLM tools and Apple Calendar on macOS. Structured access for agents, without the "please don't delete my life" risk.
Control the TIDAL desktop app from the terminal. TypeScript + Chrome DevTools Protocol, zero auth overhead — search, play, manage playlists, all pipeable.
🎧 curator
CLI-first playlist curation toolkit. Genre discovery via MusicBrainz, energy-arc arrangements, provider-agnostic. Pairs with tidal-cli:
curator discover --genre jazz | tidal playlist create "Late Night Jazz"
🚆 bahn-cli
Deutsche Bahn from the command line. Agent-first, JSON-native, cookie auth.
- Evals — shipping an agent is easy; knowing it still works next month is the hard part
- Model routing — the right model for the right subtask, not "always Opus"
- Observability — seeing what the agent actually did vs. what you hoped it did
- Composable CLIs — every tool pipe-friendly, every action verifiable
- Function before personality — it has to do the thing before it gets a voice
- Tests and quality gates are sacred — never weaken the test to make it pass
- Uniform data shapes — same type, same structure, always
- Boring architecture beats clever architecture, every time
Running. No race on the calendar right now, which means more base miles and fewer intervals.