Projects · 7 in total
Everything I've built.
The two in-depth case studies are Suma-Getsu and Tsunagi. Here is the rest, each with an honest status.
- Built & demoed live
Hojo-Grok
Built and demoed live at the Grok Bot Tokyo Hackathon, Oct 2026. Paste a link to a Japanese government subsidy programme; get a plain-language fit score, the pitfalls, and an application pack with plan, form and calendar.
Grok · LLM · web app - Working app
Event Buddy
Before an event it turns the invite into a prep checklist and a list of people worth meeting. After, it turns one sentence and your screenshots into a LinkedIn or Instagram caption. Gemini reads the screenshots.
Gemini API · vision · Flask - In daily use
Event Companion
A cross-platform to-do app for event-goers, built in a day at the teamLab hackathon (second place). Still in my daily use.
Flutter · Dart - In progress
Auria
A travel and trip-sharing app: turn saved travel videos into a shared group plan, vote on spots, share the recap. English and Japanese.
React Native · TypeScript · Code on GitHub - In progress
Savi
Personal finance for students and young workers living between countries and currencies. Offline-first, data stays on the device.
React · PWA · IndexedDB - In planning
Sol4ra
My company for AI implementation and training, plus apps, websites and marketing. Domain held at sol4ra.com; not launched yet.
Company - Studies
Data & ML studies
Survival prediction worked by hand until the mechanism was clear; a co-occurrence network of AI concepts (KH Coder, NetworkX); a self-study of sleep and screen-time data; and, in progress, a course introducing faculty to a locally hosted AI stack.
Python · NetworkX
How I build AI
Local first, then measured.
Four habits from shipping a tool a company now wants as a product, and a recorder I trust with my own conversations.
Local first
If a model can run on the device, it should. Tsunagi makes zero cloud calls once its models are downloaded.
Measure the work
Count steps and hours before and after. Suma-Getsu went from 13 steps to 5 and from about 40 hours a month to 15.4.
End to end
Model, pipeline, API, interface, and the hardware when there is some. I would rather own the whole path than hand off a notebook.
Bilingual by default
People switch between Japanese and English mid-sentence and mishear names. Systems should expect it.
