Web platform engineer and manager | TypeScript Β· Node.js Β· Next.js Β· Astro Β· Playwright Β· AWS Β· AI agents as a team
π About Β· π What I build Β· π οΈ Toolbox Β· π Current interests Β· π Activity Β· βοΈ Writing
const emmanuel: Engineer = {
name: "Emmanuel Ruiz",
role: "Software engineering manager, Arthrex (global medical device company)",
location: "Southwest Florida Β· speaks English and French",
focus: [
"π The web platform behind arthrex.com and its family of sites",
"π©» X-Ray: a Playwright quality gate that audits ~50 sites every week and on demand",
"π€ Working with AI agents as a team: skills, MCP servers, review loops, handoffs",
"π SEO, structured data and AI visibility for a regulated industry",
"π§βπ€βπ§ Leading a team that ships weekly without trading away quality",
],
stack: {
languages: ["TypeScript", "JavaScript", "Python", "Ruby", "Shell"],
web: ["Next.js", "React", "Astro", "Jekyll", "Rails", "MDX docs"],
quality: ["Playwright", "Vitest", "Lighthouse", "axe", "CodeQL"],
delivery: ["GitHub Actions", "Docker", "Kubernetes manifests", "AWS CDK", "Azure"],
ai: ["Claude Code", "MCP", "agent skills", "embeddings", "LLM-as-judge"],
},
principle: "Pace is a feature. Quality is the floor, not the ceiling.",
};X-Ray crawls a site with Playwright and checks it before it goes live and every week after: content, links, images, rendering, browser errors, SEO, accessibility, consent and brand terms. It runs on about 50 Arthrex sites and reports to a dashboard, to Teams and to Jira.
- π§ͺ 60+ checks, each a pure function over a page snapshot, every heuristic shipped with a negative test, so a false positive is treated as a bug
- π·οΈ A Site Health Score that discloses what did not run instead of awarding partial credit
- π An MCP server and a CLI, so people and agents can start an audit, read findings and ask what a check means
- π€ Built with AI agents as a team: a supervisor routes scoped tickets to cheaper models, reviewers judge diffs, hooks guard secrets and the test gate
- π The story of its first month is in One Month at the Frontier
π Internal to Arthrex, so the repository is private. The essay above is the public write-up.
| Project | What it is | Tech | Where |
|---|---|---|---|
| π₯ arthrex.com platform | The Next.js platform behind arthrex.com: structured data for videos and techniques, ISR, 404 handling, SEO hardening and performance work. | arthrex.com Β· π private repo | |
| π Astro + design tokens | Rebuilding a marketing site on Astro with a token-based design system, measured against the live site with X-Ray and Lighthouse. | π private proof of concept | |
| π₯οΈ X-Ray portal | A Next.js front end over X-Ray's reports: history per site, findings by rule across the portfolio, signed report links, deployed with AWS CDK. | π private repo | |
| βοΈ mister99.dev | Essays on building software with AI agents, published with Jekyll and GitHub Pages, with generated link-preview cards. | Visit Β· Repo |
| Project | Contribution | Tech | Where |
|---|---|---|---|
| π§± Instatic | Evaluating an open-source visual CMS for static sites; fixing what I hit on the way: CSS custom properties dropped by the publisher, imported color and font tokens whose var() references were not rewritten. |
Pull requests Β· Issues |
| Area | What I reach for |
|---|---|
| Web platforms | Next.js, React, Astro, Jekyll, Rails, Webflow, headless CMSs (Kontent.ai, Builder.io, Instatic) |
| Quality and testing | Playwright, Vitest, Lighthouse, axe, CrUX field data, CodeQL, test-first checks with negative cases |
| SEO and content | Structured data (JSON-LD), hreflang, canonicals, Core Web Vitals, brand and claim validation, AI visibility |
| Delivery and infra | GitHub Actions, Docker, Kubernetes manifests, AWS CDK and App Runner, Azure, GitHub Packages |
| AI tooling | Claude Code skills and hooks, MCP servers, multi-model supervision, LLM-as-judge evals, embeddings |
| Automation | Power Automate and Teams flows, Jira and Confluence integrations, scheduled audits |
- π€ AI agents as a team, not a tool. Supervisor and worker roles on different models, skills that encode how a repo is judged, hooks that block secrets and loose files, handoffs between sessions. The question I keep asking: how far can a small team's pace go without the quality floor moving?
- π‘ AI visibility. How LLM answer engines read, cite and rank sites, and which page-level signals move that for a regulated industry. Measured, not guessed.
- π§± Static-first publishing. Astro, design tokens and an open-source visual CMS as an alternative to Webflow for marketing sites, judged by the same audit as the sites they would replace.
- π§ Embeddings in quality checks. Paraphrase detection for approved claims, portfolio overlap and prompt answerability, on small open models.
- πͺ Consent that actually works. Checking that each consent category only activates its own tags, on real banners, with a real browser.
- βοΈ Writing it down. One essay a month on what building at this pace teaches, at mister99.dev.
Most of my commits land in private Arthrex repositories: about 1,900 contributions across 24 repositories in the past year (October 2026). The graph above counts them; the public repository list does not.
One Month at the Frontier What thirty days of building X-Ray taught me about pace, quality and leading with AI





