From models to apps,
all by my own hand.

A medical LLM that lives inside the hospital. Discharge-summary drafting, AI voice charting, referral-letter OCR. Patient data never leaves the building — fully offline.

Speech recognition tuned for medical jargon, with an AI post-processing step that fixes mistranscribed terms. Shipped as a Python SDK.

Designed and trained from the acoustic model down to the vocoder. A field only a handful of companies in Japan attempt in-house — I led it as the lead developer of a four-person team.

Reads manga speech bubbles with 115M parameters. On Japanese it beats the teacher model, and stays competitive with models 8× its size at the bubble level.

A pay-once app that shrinks images by up to 90%. I wrote the PNG and PDF compression myself in Rust and beat the commercial tools. Reached five countries on a zero marketing budget.

A desktop editor that opens AI-written HTML and Markdown and lets you fix them in place, as-you-see-them. Bundles CSS and images into a single file. Offline-first.

A Claude Skill that turns AI-written Japanese back into human prose — shipped together with the analysis behind it: "the tell of AI writing is the absence of a writer."

Invisible in the timeline, revealed on tap. It makes the "tap-to-reveal" hidden images that went around on X easy for anyone to make.

A for-fun beauty scorer that rates a face on 10 axes from a single photo, with a personal-color diagnosis. Entertainment only.

Upload a picture and it gets composited into "sublimation" artwork — a small pun-app about channeling urges into art.

Splits a tall image into a 16:9 set sized for X — a little tool for when you want it to show up big in the timeline.

"An electronic record where AI runs by default, open source." An EMR rebuilt from scratch by a physician — FHIR-native, billing included — developed fully in the open.