ClickHouse, Python, Linux.
Things I've built. Most are open source.
An intent engineering platform — an AI harness that conveys what you ultimately want to your AI on every task.
The premise: most AI failures aren’t capability failures, they’re context failures. The model doesn’t know what you actually want — not for this prompt, but in general. LifeOS carries that intent across every task, so you stop re-explaining yourself.
A folder-based task system for Claude Code — guarded locks, session logs, and an honest engineering record across sessions.
Long-running engineering work doesn’t fit in one context window. Kommander gives each task its own folder — definition, daily session logs, artifacts — plus guarded locks so concurrent agent sessions can’t corrupt the same task. Fork, switch, park, complete, archive.
Direct browser control over CDP — screenshots and coordinate clicks instead of selector archaeology.
Most browser automation fights the DOM. This one doesn’t: take a screenshot, read the pixel, dispatch a mouse event. Hit-testing happens in Chrome’s browser process, so clicks pass through iframes, shadow DOM, and cross-origin boundaries without special cases. Connects to your already-running Chrome rather than launching its own.
A curated list of Claude Code playbooks — reusable directive sets that shape how the agent works across sessions.
An index of playbooks worth installing, and a place to figure out what the good ones have in common.
Tracking what's actually shipping in AI tooling, separated from what's being announced.