AnIT

From a national contest in junior high to my research today.

  1. Junior high, year 2Failed at the national singing contestWe won the top prize in the regional round, and I went into the national round thinking we'd be fine without pushing our practice further. We failed. No one on the team blamed me, and that kindness hit the hardest. That was when I decided never to ease off when things are going well.
  2. High school, year 3Presented research in English at an international science fair, Excellence AwardAt the fair, I presented a keyboard layout designed for Japanese. I weighted keys by the home row and easy-to-reach positions, rearranged them, and measured again and again, raising the home-row typing rate from 35.8% to 81.5%. I also hosted the event two years in a row.
  3. University, year 1Started an information science degreeAt university I built a neural network in a team of five. When we split over whether to write everything ourselves instead of using a library, I took on the role of sorting it out. We learned the internals in reading sessions, used a library for the implementation, and finished it together.
  4. University, year 2Moved into a dorm with no other Japanese studentsI lived there from October 2025 to March 2026. Being with other Japanese students would have been easier, but I didn't think it would be interesting. I started conversations in the shared kitchen every day.
  5. University, year 3PdM internship at a tech companyFour of us who had never met designed a new service together. At first I only offered conclusions and we didn't click. Partway through I started sharing just the facts, left the decisions to the team, and paired every critique with an alternative. The points the judges raised with other teams were ones we had already dealt with.
  6. University, year 3Released Nihongo, a Japanese-writing skill for AII built a skill that brings AI-written Japanese closer to how people actually write, and published it on GitHub. The announcements on X have passed 5,000 bookmarks in total, and it has over 200 stars on GitHub. It has also received 3 pull requests from outside contributors.

Research

AI that warns before an accident happens, for monitoring cameras

I study AI that predicts accidents such as falls from video before they happen. I started because someone close to me has a disability, and I want to lighten the load on caregivers.

Digging in, I found that with the common way of testing, the AI can score well without looking at how people move. So now I re-test it on whether it can tell the moment just before an accident from a little earlier in the same video.

The AI has to guess which clip is right before the accident

The usual test

A video with a fall
A different video, no fall

The framing differs, so it can guess right without watching the movement

My test

Same video, 3 s before
Same video, 1 s before

The framing is the same, so only the warning signs give the answer

ToolsPython, PyTorch, VideoMAE, V-JEPA 2, YOLO11, LLaVA and more

More pages