I'm a doctor in the imaging department, mostly. Now and then I build a small tool, or train a model on medical images. Small work, done slowly, and checked against the evidence.

A terminal-native downloader built on yt-dlp and rendered with Bubble Tea and Lip Gloss. Concurrent downloads, per-task progress locked to a fixed width so the numbers never judder, a bilingual interface, native notifications, and a failure log that actually respects your time. Go · MIT

A single-file Ghostty config for macOS: JetBrainsMono Nerd Font with a PingFang SC fallback for clean CJK rendering, a theme that follows the system, a translucent window, and iTerm2-flavored keybindings. The terminal is where I live; I'd rather not live somewhere shabby. Maintained with CI that validates the config on every change. Ghostty · MIT

Things I recently took from idea to merge — branch, review, CI, homing in on the details, then a clean squash. Every step green before it moved on.

• ghostty-config #2 — feat: default working directory for new windows. +validate-config clean on both macOS runners, then merged.
• ghostty-config #3 — docs: synced README & CHANGELOG to match; markdownlint, zero warnings.
• ghostty-config v0.2.0 — cut a release whose notes came out of the CHANGELOG, no drift.
• Workflow discipline: branch → PR → CI green → squash → prune, every time. Config + docs ship in one atomic PR.

Reviews offered upstream — code handed back with the polish applied, not just a verdict:

• Project-MONAI/MONAI #9134 — approved a MeanIoU ignore_index fix after reproducing the old channel-wipe path (class-0 IoU ≈0.67 on the ignored voxel → 1.0 after spatial masking, matching Dice); suggested an include_background=False assertion to lock the mask-vs-channel interaction.
• averygan/reclip #79 — reviewed a yt-dlp -- separator-only injection fix against yt-dlp 2026.08.19; confirmed all three call sites, and pointed out it is a weaker subset of #58 (no is_safe_url early reject).
• spotDL/spotify-downloader #2795 — reviewed a YouTube Music locale-fallback retry (en/de client cycling); flagged the silent de→en default flip, the duplicated third attempt, and the missing mock regression test.

• Write the smallest thing that solves the problem, then delete what remains.
• Badges are welcome; declarations of "impact" belong in journals.
• If a tool needs a tutorial, it isn't finished.