AI systems, engineering, execution.
Nine years building practical AI across Japan and India with a bias for clean execution. Today my focus is GenAI: agentic LLM pipelines, RAG, and fine-tuning, built on a deep computer-vision and synthetic-data foundation.
I work where technical depth meets cross-cultural delivery.
Six years in Japan taught me that the hardest part of shipping ML isn't the model. It's the conversation between the people who define the problem and the people who solve it.
From Level-4 autonomous trucks at Hino, to social robots deployed at 10+ research institutions through Honda, to computer-vision and ML systems at Sony, I've spent my career closing that gap. Building from scratch, leading small teams, and translating ambiguity into things that work.
More on the about page, or skip ahead to the resume.
and India
using work I shipped
behind one eval set
owned end to end
Things I've built that actually shipped.
Haru, the social robot platform
Led three applications on Honda's Haru robot: an emotionally-intelligent Telegram remote, an international classroom mediator (Japan ↔ Australia, published in research papers), and an NLP collaborative storyteller. Built the virtual simulation world from scratch.
CAD symbol detection for construction
Led a 3-person team automating cost estimation from architectural blueprints. Solved data scarcity by generating synthetic training data with Blender, Unreal Engine, and OpenCV. No manual annotation required.
Agentic LLM pipeline for structured decisions
Architected a pipeline that turns audio into structured, cited output: fine-tuned transcription feeding a RAG-grounded discussion between specialist agents, with a schema that fails loudly rather than inventing.
emotion-english, an NLP model in the wild
A text-classification model for emotion detection in English (20 classes). Trained, published on Hugging Face, and quietly running in open-source projects around the world.
Things I write, ship, and tinker with.
Japanese Compass, a one-person GenAI product
A Japanese-learning brand where an automated content factory turns lesson JSON into carousels, reels, quizzes, audio, and video, then publishes on schedule via the Meta Graph API on a Cloudflare Worker cron. Site, pipeline, and video generation, all built and run solo.
Notes on engineering, philosophy, and building products
Long-form thinking on the systems side of ML. How models meet real workflows, how Japanese and Indian engineering cultures differ, and what I learn while building.
Lab, interactive prototypes and experiments
Live technical proofs of concept. Things I build to understand a problem, test an idea, or just because the question wouldn't leave me alone.
Open NLP models on Hugging Face
Published models including emotion-english for text emotion
classification. Trained, shared, and running quietly across the open-source
ML community.
Open-source Python tools
secscan-tool, an automated web security scanner with
AI-tailored remediation, plus printj, pyjeasy,
and jaitool. Small, sharp, and built from the friction
of daily work.