Autoware evaluation lab
A full Linux build of Autoware run by the official procedure, completed in 6h35m26s with all 621 target packages succeeding. The environment, the inputs and every failure along the way are recorded.
How is new technology invented—and brought into society?
Projects I built
Anomaly-detection pipelines, model quantization, an Autoware build, and browser tools. Explore the implementation, how to run it, and the measured results.
IMS bearing anomaly detection
I built an anomaly-detection pipeline from bearing vibration data to model training and alarm evaluation. On the public IMS dataset, the fixed protocol that holds out entire test runs did not establish an advantage for deep autoencoders.
Vibration Anomaly Lab
I built a synthetic vibration generator and an anomaly-detection experiment with feature extraction and model evaluation. Scores fell when machines were kept separate between training and evaluation, exposing results inflated by data leakage.
mct-bench
I built a CLI that connects model quantization, ONNX export and CPU measurement. On 10,000 Fashion-MNIST test images, W4A8 accuracy fell from 94.76% to 94.22%; the exported ONNX file grew by about 0.18%.
KAMIWAZA Gate
I built a document-checking tool with a Python library, CLI and review interface. YAML rules identify issues; a person can approve an exception or return a document, and the audit history preserves both the finding and the decision.
Patchlight
A small browser workshop for connecting input, process, and output blocks and exploring their effects through light, sound, and color.
Nori
I built clipboard-history apps for macOS and Windows. Text, images and files retain their own previews and actions, while hints in the interface show how modifier keys and permissions affect each action.
Vesuvius Challenge
A patch and measured comparison for official tools that silently selected the CPU on Apple Silicon even when MPS was available.
UGUIS
A small iOS work about passing time, with a rising and falling tide and a bush warbler that calls when you pause.
Interactive works
DOCUMENT → APPSPEC → APPLICATION
KAMIWAZA →
A photograph of a form becomes a working application.
AI HACKATHON 2026 TOKYO
Finalist · 56 teams
- TypeScript
- React
- Next.js
How it is built
- AppSpec / JSON
- Describes the fields, tables and records read from a form.
- React
- Renders that specification as an editable application.
- TypeScript / Next.js
- Defines the data types and connects the screen to the analysis API.
CELL CITY →
A simulation of Shibuya, running in your browser.
- JavaScript
- Three.js
- Web Workers
How it is built
- Three.js / WebGL
- Draws the buildings and roads in 3D.
- Web Workers
- Builds parts of the city off the main thread.
- TypedArray / cache
- Holds geometry within a chosen memory budget.
GENESIS →
Small rules create a world of feeding, birth, and change.
- JavaScript
- WebGL2
- Web Workers
How it is built
- Web Workers
- Advances the ecosystem separately from drawing it.
- WebGL2 / Canvas 2D
- Draws the organisms, with a Canvas fallback.
- Seed / snapshot
- Saves the simulation state so the same run can continue.
Art for Well-being
彩 / Irodori ↗
A playful mochi made with quiet hours in a hospital room in mind. Press and release with touch or a mouse.
- Canvas 2D
- Web Audio
- Web Bluetooth
How it is built
- Canvas 2D
- Draws the mochi and its response to pressing and releasing.
- Web Audio
- Shapes the sounds that accompany the interaction.
- Web Bluetooth / MESH
- Connects physical blocks in supported browsers. Touch and mouse also work.
Work in progress
Haptics stack
Separating haptics into a core, operations, engine integration, and a test environment.
Reproducible autonomous driving environment
Building an open-source driving stack toward repeatable runs, with build and evaluation records.
Recognition
- StartupBase U18 — 1st place, team
- Ministry of Foreign Affairs scholarship — National Taiwan University
- British Government scholarship — The University of Manchester
- IT project management simulation — 3rd place, team
Project management using waterfall and methods informed by distributed agile.
- Quality Kaizen (QCC) — 1st place in the company, team · Presented company-wide
Over three months, built a RAG system that retrieves domain-specific knowledge, combining calculations based on user-supplied conditions with qualitative reports to improve the quality of the information provided.
- Internal coding test — 1st place among around 200 participants, individual
Completed coding tasks focused primarily on frontend languages.
- Manufacturing anomaly detection — special award, team
Applied LightGBM and other methods to anomaly detection in manufacturing data.
- AI HACKATHON 2026 TOKYO — finalist, individual · 56 teams
Over approximately two months, worked from concept development, interviews and problem definition through requirements alignment and implementation. Built deterministic rendering that turns generated specifications into interfaces using fixed rules.
- Japanese localization — specialized training applicationOfficial contributor; credited by the developer.
Over approximately a year and a half, alongside Japanese localization, worked on proof-of-concept validation, stakeholder coordination and application improvements with engineers abroad.
About
How is new technology invented—and brought into society? I build things, try them, and bring what I learn back into the next design. Through that cycle, I explore how technology becomes something people can use. What matters to me is innovation: creating new ways of doing things and making them work in society. The projects here record what I tried, what I could verify, and the questions that remain.