I’m a product engineer working where product experience meets engineering correctness. I build small, local-first tools for agent workflows, then make their behavior explicit enough for people to inspect, understand, and trust.
A local-only, read-only panel for seeing what Codex, Claude Code, and OpenCode can actually do on your machine—without exposing prompts, commands, credentials, or transcripts.
A local-first trace observatory that follows one coding-agent turn from observable context through tool execution to the final response—without pretending missing evidence exists.
- SecondEar gives agents timestamped text from public Bilibili videos, preferring native subtitles and falling back to local transcription.
- Palbreed finds deterministic shortest breeding routes offline, with explicit constraints and machine-readable results.
- kimiusage turns local Kimi Code session logs into offline usage reports with estimated costs.
- ax — tighter CLI contracts across output modes, numeric flags, authenticated extraction, redirects, media types, and JSON continuation.
- Vercel Labs / Skills — full-depth skill updates and new agent integrations.
- Mindwalk — Agent Lens, Codex event compatibility, and faithful UTF-8 errors.
The recurring theme is small seams with outsized consequences: ambiguous flags, hidden state, lossy traces, and interfaces that imply more certainty than the evidence supports.
Game soundtracks, books about products and economies, Palworld, and Wuthering Waves.