Projects

Selected and funded by the AKATSUKI Project, a Japanese government program for young IT creators. A Chrome extension (MV3) for Google Play Books that supports reading with hints and on-device quizzes (Gemini Nano), then withdraws that support too gradually for the reader to notice, driven by a controller over in-session reading events.

Privacy and design rules are enforced in code — the page is never modified, reading data stays on the device and the book's text never leaves it — and automated tests check them on every pull request.

Built a production crash-triage tool for ChromeOS engineers, using Gemini CLI to automate the analysis of kernel crashes that had been investigated by hand, backed by a Python + SQL pipeline that turns noisy production crash logs into structured context. Reduced engineer analysis time by 92% on real production crash data, while giving the model only what it needed: approved environments only, only the fields it needed, and log contents treated as untrusted input.

Solo-built app: crop a region of the map and send only the places inside it as a time-limited link, with no login needed to view. Next.js web and React Native iOS/Android on one Supabase/PostGIS backend, with native Swift/Kotlin share extensions so spots can be added from any app's share sheet. Live on the App Store.

Solo-built Flutter app released on App Store and Google Play. Search songs via Apple Music API, tag with key and mood, get a random pick based on how you feel. Backend: FastAPI on Google Cloud Run.

As part of a 7-person team, tackled the SSVEP universality problem — why some brains are harder to decode than others. I developed the channel selection and evaluation pipelines and built the comparison harness; the team's approach improved accuracy by +22.5pp on hard-to-decode subjects, reaching 95% peak accuracy.

Showed that a brain-to-image reconstruction pipeline still produces plausible, category-consistent images from shuffled or random fMRI, and that statistics which do not use the true stimulus, such as the norm of the decoded features, cannot tell the difference.

Brain Control (BC) compares the variance of the decoded features with the brain–stimulus pairing kept and broken: BC = 1 means no brain information. It carries the same information as the ANOVA F statistic, but unlike F it does not grow with the number of trials, so it can be compared across experiments as an effect size. Replicated in all 5 subjects (p = 0.0017). Paper and reference implementation on GitHub.