shunzz.com
Spec-driven × AI agents × full-stack delivery. I don't just use AI to write code — I manage how AI works, and I build tools for it.
01
Japanese Literature major turned AI-Native Software Engineer (JLPT N1). I ship enterprise systems (Text-to-SQL, an SRE triage agent, a job scheduler), ML models and cloud infrastructure with a spec-driven, AI-agent workflow, and launch my own products with the same workflow in days; I write Claude Skills, built an MCP Gateway, and teach the workflow internally. In the AI era an engineer's value is judgment and process design, not typing speed.
Japanese Literature degree, JLPT N1. My career began with language; a six-month AI/cloud bootcamp moved me into engineering. I built AI applications at a startup (content-to-video, LINE avatars and voice cloning) and now work as a Senior Systems Analyst at a retail group's IT subsidiary, spanning data engineering, DevOps and cloud architecture.
Over the past year I moved my workflow to spec-driven development with AI agents: brainstorm → design spec → implementation plan → TDD, where AI executes and I review and decide. Real projects grew it into my own version — design phases produce only documents, and subagents implement only after decisions are final; a self-made dispatch skill routes each task to the cheapest model that is good enough; one shared AGENTS.md drives two agents; an automated AI code-review loop. With it I independently delivered an enterprise allocation platform (2.5 months), a CDP's AI service and job scheduler (88 commits and 1,190 unit tests in one day), took Text-to-SQL from PoC to a production service, and built a company MCP Gateway in 3 days. After hours, same workflow: a ramen map live on day one, a Japanese typing trainer with 470 tests in 3 working days. Every number is in the git history.
I don't just use AI tools — I package team domain knowledge as Claude Skills with evals, built the gateway that lets every AI tool in the company reach internal APIs safely, wrote a Claude Code mod that shows context usage and cost in real time, and draw boundaries for AI agents: proposal-only changes gated by automated tests, zero shell and zero cloud permissions, never writing to third-party platforms on a user's behalf. I also learned to reject upgrades with data: a 65-case eval showed the bottleneck was the rules, not the model, so I built human-precedent override instead and triage accuracy went from 12% to 80%. Next: build something that actually matters.
02
03
AI job-search assistant: upload résumé → AI diagnosis → auto-crawl listings → fit ranking → tailored cover letters → application board
Built so others could solve their own problem with a tool. The launch post on Threads passed 10K views and it has real users. It later grew a local edition: reads your logged-in job-site data and runs a tool-use chat agent as a "job-search commander" — but the agent never applies on your behalf.
Azure Data Factory failure triage: read-only against ADF, detects failures every 5 minutes → de-duplicates into incidents → rules first, LLM as fallback → a Teams card tells the on-call engineer the conclusion, what to do now and what to check first
Rejected a model upgrade with data: on 65 eval cases Opus looked far ahead, but with duplicates removed it tied with gpt-mini — the bottleneck was the actions the rules suggested, not the model. So I built human-precedent override instead. The learning agent is proposal-only: a human clicks apply, pytest reverts failures, passing changes commit; zero shell, zero Azure permissions. Designed and built solo in about five weeks.
Natural-language queries over 920M transaction rows: understand → code-side guards → option-based clarification → generate SQL → validation and an EXPLAIN gate → code computes statistics, the LLM only interprets
From PoC to a production service: sqlglot dialect packs for both T-SQL and PostgreSQL, a semantic model where every fact carries confidence and evidence, 10 ADRs. Guard rule: any condition that cannot become SQL is asked back, never silently dropped. Once other systems asked to onboard, the positioning became "the execution engine plus a 13-step onboarding process is the product".
Taipei ramen map: daily Google Maps harvesting → open-now / metro / style filters → a "where today" dice → ramen diary and friends
Fourteen minutes from pricing Google Places API to dropping it: the scraper runs at home and the cloud only reads. Tests showed datacenter IPs get degraded, stripped-down pages, so the collector stays home. LLM classification of 491 shops cost US$2.5. Worker-rendered SEO pages, self-written Google OAuth, $0 a month.
Japanese typing trainer: romaji input judged key by key into kana, from kana charts to N1 vocabulary, timed leaderboards, and your own lyrics synced to YouTube
Seven worktree subagents built it in parallel and merged the same day. The server recomputes every score from the client's key log and applies six anticheat rules — it never trusts client numbers. Lyrics copyright was researched (JASRAC, Taiwan copyright law) before design: shared timelines store only hashes, notice-and-takedown, three strikes.
Housing decision tool: enter an address → commute times (bus/metro + scooter/car road graph) → rent and sale price benchmarks → mortgage and monthly cost → hazards, crime, amenities → favorites and comparison
The first plan was to scrape listing sites. After reviewing fair-trade rulings I judged the legal risk too high: users bring their own listings and the product sells only the analysis layer built on commercially usable open data. Commute times run my own Dijkstra on the OSM road graph; Zod schemas are shared across client and server, and Vitest runs in both workerd and Node.
Wraps internal enterprise APIs as MCP tools so any AI coding agent reaches them safely through one entry point
Design focus: per-client auth, SSRF protection, encrypted upstream credentials, audit logs with sensitive-parameter masking — both input and output are recorded. Tool catalog lives in a DB; approved tools go live without code changes or restarts. One admin page generates connection configs for eight AI clients.
Retail allocation platform, CDP batch evolution and job scheduler, performance root-causing, demand-forecasting ML, Azure IaC
Delivered solo with AI collaboration at a retail group's IT subsidiary: a 2.5-month allocation platform (FastAPI + React 19); ADF polling replaced by a queue, then by a single-table scheduler, with the production cutover led by me; three measurement-first performance root-causes with zero builder changes; LightGBM/CatBoost allocation models; Bicep IaC reconciled with portal drift.
04