Building end-to-end systems — from statistical modeling to shipped apps.

Data/ML engineer at a Japanese e-commerce company. I work across the whole line: designing the statistical method, building the pipeline, wrapping it in an API, and shipping the app that sits on top of it.

• 📊 Survival analysis & causal inference — propensity score matching, Kaplan–Meier, Cox proportional hazards
• 🤖 LLM applications — RAG systems, agent pipelines, vector search
• 🛠️ Automation — turning manual, repetitive work into pipelines that run themselves
• ✍️ I write about what I build at subbrain-works.net

Languages

Data & ML

LLM & Retrieval

Cloud & Infra

Tooling

Most of these live in private repositories — clinical data and unreleased products. Happy to walk through the design and code in a conversation.

A reproducible pipeline for time-to-event analysis on clinical data, built for a research collaboration in the healthcare domain.

• Confounder adjustment via propensity score matching, with balance diagnostics
• Kaplan–Meier survival curves and Cox proportional hazards modeling
• Packaged as a reusable src/ library (preprocessing / matching / survival / evaluation / visualization) so that every result is reproducible from raw data

Python lifelines scikit-survival statsmodels pandas

A multi-tenant SaaS that turns any WordPress site into a RAG-backed chat experience. Site owners pick pages from the WP admin screen; an async pipeline crawls, embeds, and indexes them, and a shortcode drops the chat widget onto the live site.

• Firestore Vector Store for retrieval — near-zero fixed cost, real-time updates, native GCP integration
• FastAPI on Cloud Run backend with per-tenant API key auth
• React + TypeScript + Vite + Tailwind console for RAG parameter tuning, knowledge management, and answer previews
• Fully reproducible locally via the Firestore Emulator, with CI on GitHub Actions

FastAPI GCP Firestore React TypeScript PHP

An end-to-end pipeline that generates long-form Lo-Fi music videos and publishes them to YouTube without human intervention — a production system, not a demo.

• Audio generation with diffusion models, stitched into hour-long tracks
• Video composition and rendering via MoviePy / FFmpeg
• Automated upload, metadata, and scheduling through the YouTube Data API

Python diffusers MoviePy YouTube Data API

• Life is Quest — a gamified task manager that reframes goals as RPG quests, with a retro-game UI (Swift / SwiftUI)
• AWS CLF Exam App — a certification study app with LLM-generated practice questions, a Firestore-backed Python service, and progress analytics (Swift / Python / Firestore / OpenAI API)
• Python Analysis Template — an opinionated starting point for analysis projects: uv + src layout + Ruff + mypy + pytest, so a new project is reproducible from day one

_{Language and commit stats include private repositories — repository names and commit contents stay hidden.}

I publish notes on data analysis, LLM applications, and automation at subbrain-works.net — mostly write-ups of the problems I hit while building the projects above.