Ryusei (Ray) Takiya

For recruiters

Work authorization

IEC Working Holiday approved: open work permit for 12 months from December 2026, usable twice. No sponsorship needed to start; once I’ve proven my value, I’d welcome visa support to stay longer.

Start

December 2026

Location

Relocating to Vancouver, BC in December 2026; open to anywhere in Canada. On-site, hybrid or remote.

Employment

Any type: full-time, part-time or contract

Experience

4+ years of professional software engineering

Training

Programming school, 7 months (2022), before starting work

Languages

Japanese (native); English: written work in English on GitHub and Slack is fine; spoken English is elementary and I’m studying (3-month intensive course in the Philippines)

Education

High school diploma; programming school, 7 months (2022)

Experience

Highlights

• Automated ad submissions that used to be done by hand, building a system on the Google Ads API with an AI regulation check built in. It works with separately managed systems for ad generation and for billing and operations; I own the submission automation service. It cut a client’s monthly work from 80 hours to 5 (about 94% less, measured by the client before and after rollout); the client thanked me by name. AWS and GCP are joined with Workload Identity Federation: the ECS task’s IAM role authenticates as a GCP service account to reach the Google Ads API, so no long-lived keys or refresh tokens are stored. I also split YouTube uploading out of the submission system into an independent service with its own infrastructure and database, sharing the uploaded URLs with the submission side through a data integration service
• Own the whole lifecycle on my own: requirements, design, build, test, release and operations
• Built an AI-agent system that automates software development and rolled it out to the team: harness and loop design plus AI reviewers automate reviews, tests and deploys, and about 35% of PRs now merge automatically
• Cloud on AWS and GCP, with infrastructure as code in Terraform
• Added Google account sign-in (OAuth 2.0 / OpenID Connect) with Amazon Cognito; built Sign in with Apple and Google into several of my own products
• In my own products, AI features that read text from photos and transcribe speech into shopping-list items (Shopping Rhythm)
• Led a team of five
• Worked in a large team of 100+ engineers
• Maintained a POS system with 3M+ members and hundreds of millions of records a year, integrated via APIs with an e-money system on a separate platform
• Designed and built from scratch a multi-tenant software package for the public sector

Client names and details are confidential.

Skills

AI & agents

Claude Code, Codex, Multi-model review, LLM evaluation

Backend

Java, Spring, Go, Python (FastAPI), MySQL

Web & frontend

TypeScript, React, Next.js

Cloud & infrastructure

AWS (ECS), GCP, Docker, Terraform, GitHub Actions (CI/CD)

Security & auth

OAuth 2.0 / OpenID Connect (Amazon Cognito, Supabase Auth), Workload Identity Federation (AWS ↔ GCP)

Mobile · PlacePouch, Shopping Rhythm

Flutter, iOS & Android, 13-language localization

Backend · PlacePouch

Supabase, Postgres, TypeScript

Products (built and run solo)

• PlacePouch — Save the places you want to go, scattered across social media, all in one place. See them on a map and share them for outings and trip plans.
• Shopping Rhythm — Recurring items and everyday shopping in one place Share with your family so nothing is forgotten
• PlacePouch, built and run solo: #1 Travel, Japan App Store · #11 All apps, Japan App Store · 10,000 installs on day one · ~30,000 installs by Sep 2026

Open source

• flutter/packages #12998 (Approved by the Flutter team, waiting to be merged) — Found and fixed a bug in image_picker_ios where images couldn’t be saved.
• master-llm (Paused) — Built an evaluation harness that measures accuracy, cloud cost and time for each task, comparing a local LLM against Claude alone. On top of it I designed a router that sends each task to the cheaper model when it can: tasks become embedding vectors, and similar past tasks (kNN) show which model succeeded. Development is paused before the router was built.

How I ship with AI agents

• Issue — A person writes the requirements, goal, acceptance criteria and what is out of scope.
• Plan — The agent reads the issue, designs the feature and tests, and decides the order of work.
• Build — Claude Code writes the tests first, then the code (TDD).
• Verify — Tests, static checks and screen checks run; results are kept as evidence for each acceptance criterion.
• Review — In a separate session, Claude and Codex review the diff.
• Pull request & merge — The PR is split by risk: