Software engineer — Tokyo, Japan
What have you actually shipped?
AI-focused software engineer in Tokyo. Two years building RAG and LLM systems for Japanese public-sector and construction clients.
From a remote internship to on-site delivery for Japanese public-sector clients, in about three years.
JUL 2023 – JAN 2024
JAN – JUN 2024
JAN 2024 – JAN 2025
FEB – MAY 2025
JUN 2025 – PRESENT
What I have shipped with, grouped by what it is for. No self-rated percentages — the projects are the evidence.
Six builds, led by the ones with numbers attached. Most are client work delivered on-site in Tokyo.
Pacific Consultants 株式会社
LangChain and LLMs over roughly 400,000 construction and tender documents, covering search, summarisation and Q&A. Out of the sandbox and into daily use in about two months; retrieval tuning alone took roughly 1,400 hours of manual document work off the year. Python and Express behind it, React and TypeScript in front, Docker out to DigitalOcean with Cloudflare ahead of it.
“Most of the accuracy work turned out to be retrieval, not prompting — chunking strategy, hybrid search over pgvector, and a reranking pass, measured against an eval set built from real user questions.”
MLIT / 国総研
Staff describe what they need in a chat, and the system assembles an outline, body text and slides. Deck preparation went from days to under an hour. Existing documents are referenced automatically so the substance stays consistent between authors. Azure Web Apps and Functions, with Application Gateway and a virtual network restricting the public route.
Otaru Canal · Kashiwa City
LLM-driven querying and visualisation of geospatial data, built to address overtourism congestion. Transit, rail, ferry, event, weather and retail data merged and normalised into 250m mesh cells, with congestion forecasting on top. MapLibre for the map, PostGIS on Azure and Neon, QGIS for preparation.
Automated extraction from government tender PDFs into structured CSV and Excel. PyMuPDF for parsing, multithreading for throughput, and region-aware rules for the long tail of format differences between prefectures.
Spaced-repetition learning SaaS
A subscription learning app built to production standards: a scheduling engine, RAG-backed generation of personalised mnemonics, and progress tracking. Next.js on the front with a separate FastAPI service doing the AI work, PostgreSQL and pgvector for retrieval, Redis for sessions and rate limiting, Stripe for billing.
Conversational assistant with token-level streaming, persistent multi-session history and per-user rate limiting. Streaming alone cut perceived response time by around 40 percent. A prompt-template layer keeps conversation context inside the model's token budget.
Open to permanent in-house engineering roles in Tokyo from December 2026.
Resume and 職務経歴書 available on request, along with a 履歴書 if your process needs one.
