Senior Full-Stack Engineer
Autify, Inc. · Tokyo, Japan
April 2022 – Present
Test automation SaaS, 300+ customer companies.
- Maintained the web / mobile test automation pipeline serving 300+ companies; reduced mobile device wait time by ~50% by pre-warming devices before test execution
- Built, as sole author, a fleet-wide dependency-update control plane where a human always merges: self-hosted Renovate as a GitHub App, deterministic rules first, then headless Claude Code grades each bot PR's risk and can auto-fix it on opt-in, sandboxed on per-repo least-privilege App tokens in separate locked-down non-root containers; grown to 16 managed repositories, replacing the vendor-hosted bot
- Built an incident-response bot whose read-only access is enforced by credentials rather than tool flags, so a prompt injection can't write; it answers on-call questions from 7 read-only MCP sources behind an MCP gateway on ECS Fargate, piloted in live incidents
- Built both products' MCP servers (web and mobile), turning every public API endpoint into an MCP tool, with write guards so an editor save can't silently overwrite an agent's change and a per-token rate limit (300 requests/min)
- Replaced an image-classification model with a deterministic DOM-based resolver, shadow-tested it behind a canary (3,505 decisions, 99.5% agreement where resolved), and recommended retiring the model, which was then removed
- Run my own engineering through a gated coding-agent loop: each Claude Code session works in an isolated worktree, an independent cross-harness review runs before a PR is posted, and 16 skills and 5 hooks I wrote encode the team's PR process
Full detail
- Drove data analytics initiatives across the organization's Business Intelligence stack (cost tracking, data governance, incident reporting)
- Created an ETL pipeline (AWS Glue / Python) that processed 20M+ unstructured log records in days instead of weeks for AI self-healing analysis
- Built, as sole author, a fleet-wide dependency-update control plane: self-hosted Renovate as a GitHub App, deterministic rules first, then headless Claude Code grades each bot PR's risk and can auto-fix it on opt-in, with a human always merging; 55 of the repo's 59 merged PRs, grown to 16 managed repositories, replacing the vendor-hosted bot
- Sandboxed the pipeline's AI agents after finding that one write token could reach every managed repository: per-repo least-privilege App tokens, assess and fix in separate locked-down non-root containers, and the signing key behind a protected environment
- Built both products' MCP servers (web and mobile), turning every public API endpoint into an MCP tool; added agent scenario-authoring tools with write guards so an editor save can't silently overwrite an agent's change, and opened the mobile server to any API token behind a per-token rate limit (300 requests/min)
- Built an incident-response bot that posts context and answers on-call questions from 7 read-only MCP sources behind an MCP gateway on ECS Fargate, with read-only enforced by credentials rather than tool flags so a prompt injection can't write; piloted in live incidents, hardening its connections after each one
- Replaced an image-classification model with a deterministic DOM-based resolver, shadow-tested it behind a canary (3,505 decisions, 99.5% agreement where resolved), and recommended retiring the model, which was then removed
- Run my own engineering through a coding-agent loop: 213 of 313 PRs (Jun–Oct 2026) authored through Claude Code sessions in isolated worktrees, independent cross-harness self-review before posting, and 16 skills and 5 hooks I wrote that encode the team's PR process
- Maintained the test automation pipeline for web / mobile automation, serving 300+ companies
- Architected solutions for email delivery and PDF generation
- Debugged native Selenium WebDriver and Chromium internals to stabilize browser automation
- Active development on Chrome extensions for recording and replaying browser actions
- Reduced mobile device wait time by ~50% by pre-warming devices before test execution
- Assisted in building the test automation pipeline for a desktop application; drove development of the scheduling feature
- Deployed a self-serve infrastructure provisioning API (Pulumi) with AWS IAM permission isolation
- Prototyped a natural language AI Agent for mobile test automation (prompt adherence, element location, model selection)
- TypeScript
- React
- Redux
- Apollo GraphQL
- Electron
- Chrome Extension API
- Appium
- Ruby on Rails
- GraphQL
- Postgres
- SQLite
- Github Actions
- CircleCI
- Jest
- Testing Library
- Mock Service Worker
- RSpec
- Playwright
- AWS (ECS/EKS)
- Kubernetes
- Pulumi
- Terraform
- Holistics
- Athena
- Redshift
- Airbyte