What I've built

The same to-do app implemented across every major JavaScript framework, one honest, side-by-side reference for comparing them.

The MV* framework explosion left developers with no fair way to compare Backbone vs Angular vs Ember vs the next thing. Everyone was picking blind.

It standardized the "todo app" as how the industry teaches and compares frameworks, and it has held that role for over a decade, first for framework authors and programming communities, and now for AI labs building coding models. Its code even became the core workload for Apple's Speedometer browser benchmark, so a teaching project ended up shaping how browser engines get tuned.

One of the web's first big scaffolding tools. It put a generator layer (the yo command) on top of the modern front-end stack, pulling in Grunt for builds and Bower for packages, so a best-practice project was one command away.

Starting a modern front-end project meant hand-wiring a dozen tools. Most people either got it wrong or gave up.

We built a large community around it, and its individual pieces, especially the generator model, became the DNA of nearly every framework CLI that followed, from Create React App to Angular CLI to Nuxt.

I played a big role in Material Design's launch on the web. I wrote, co-authored, and led Material Design Lite, the first implementation of Material Design in plain, framework-free CSS, HTML, and JavaScript, and I worked directly with the Material Design team to figure out desktop specifications for a range of UX elements. Alongside it I worked with the Polymer team on Polymer and Web Components, and on the Polymer paper-elements library of Material components.

Material Design existed, but using it meant pulling in a heavy framework, and the desktop side of the spec was thin. The web needed a vanilla option and real desktop guidance.

Material Design Lite grew past 32,000 stars and its ideas graduated into Material Components for the Web. The Polymer paper elements became part of the official set of Material implementations.

I started a team around service-worker libraries. We began with sw-precache and sw-toolbox and turned that work into Workbox. In parallel we built our own web push messaging solution.

Service workers unlocked offline, caching, and push on the web, but writing them by hand was low-level and easy to get dangerously wrong.

Workbox became the canonical service-worker library on the internet, still the default anytime someone builds on service workers, and something we kept investing in for years. Our push messaging work ultimately fed into FCM, Firebase's push notifications for the web.

I played a large role in the PWA rollout, working hands-on with many top sites and brands, Twitter, Pinterest, and others, and collaborating with React, Angular, Nuxt, Next.js, and Webpack to make sure PWA support was first-class inside the frameworks themselves.

The mobile web was slow and unreliable, and users on flaky networks were being pushed toward native apps. PWAs promised app-like speed, offline, and installability on the open web.

Twitter Lite cut time-to-interactive by 50% on slow devices. Pinterest's PWA drove a 44% lift in ad revenue and 40% more time on site. The patterns fed straight back into the frameworks and Chrome's tooling.

A run of web-performance tooling and practices. I introduced performance budgets to the ecosystem and helped land budget support in webpack, then built two prefetching libraries: Quicklink (prefetches in-viewport links during the browser's idle time) and Guess.js (reads analytics at build time to predict and prefetch the next page). My Critical tool for inlining above-the-fold CSS came out of this same push.

Performance regressions crept back in silently with every release, and the next page could feel instant if prefetching were easy and safe. Teams needed both guardrails and runtime tools, not just good intentions.

Performance budgets became standard practice, wired into build pipelines like webpack so a regression can fail the build. Quicklink and Guess.js became the direct blueprint for the browser's Speculation Rules API, Chromium cited Quicklink as prior art when it built the idea into the platform.

I co-authored large parts of Speedometer 2, the cross-vendor benchmark for real-world web-app responsiveness, built with Apple's WebKit team, and later mentored the work on Speedometer 3.

The industry needed one shared, realistic way to measure browser speed on modern framework workloads, not synthetic micro-benchmarks.

It has been enormously influential in how the industry and browsers approach performance and speed up the internet. Apple cited Speedometer when it marketed the M1 chip, and every browser and silicon vendor now optimizes against it.

I played a heavy role over many years in standardizing and shaping Core Web Vitals, the user-centric performance metrics: Largest Contentful Paint, Cumulative Layout Shift, and later Interaction to Next Paint.

"Does it load fast for a real user?" had no shared, honest definition. Everyone measured performance differently and argued past each other.

Core Web Vitals became a Google Search ranking signal and the industry's shared language for performance. Site speed went from a niche developer concern to a boardroom one.

Through Chrome Aurora, my team shipped performance directly into the frameworks developers already use: the revamped next/image in Next.js (with Vercel), the Angular image directive, and font optimization for Nuxt, among many others.

Asking millions of individual developers to become performance experts will never scale. The leverage is to fix it once, at the framework level, so everyone downstream gets it for free.

Real, measured wins: the Angular image directive cut Land's End's median LCP from 12s to 3s, and the Next.js image work improved partner LCP by up to 60%. Good performance became a side effect of building for the web.

The other half of that work, through the Galaxy effort: engineers contributing directly to WordPress, and consulting with Wix and Elementor, to improve Core Web Vitals across the CMS ecosystem.

A huge share of the web runs on a handful of CMSs. Improving them lifts millions of sites at once, including all the ones whose owners will never read a performance guide.

CMS stacks got measurably faster at the source. WordPress alone, which powers a large fraction of the web, shipped defaults like native lazy-loading and fetchpriority on its main image, moving Core Web Vitals for millions of sites in a single release.

I had heavy input into Priority Hints and co-authored the specification, as a named editor, that became the fetchpriority attribute, giving developers the ability to steer the priority of individual network requests.

Developers could preload a resource, but they couldn't tell the browser which resources actually mattered most. There was no steering wheel.

It became the de facto way people steer loading on top of things like preload, and shipped in every major browser. Google Flights cut LCP from 2.6s to 1.9s with it; WordPress later turned it on by default.

Quicklink and Guess.js proved that idle-time, viewport-aware, predictive prefetching worked. I then consulted heavily with the team bringing those ideas into the browser, providing regular input as they became the Speculation Rules API for prefetch and prerender.

Predictive prefetching worked beautifully in libraries, but it belonged in the platform, where it could be safe, native, and universal.

It shipped as a web standard in Chrome and is now used in production by Google Search, WordPress, and Cloudflare for near-instant navigations. Chromium cited Quicklink as prior art.

I led two evolutions of DevTools. First, the org's shift into post-load performance: user-journey and user-flow analysis, the Core Web Vitals Chrome extension, the Recorder for capturing and replaying flows, and scripting. Then the push into a human and AI future, DevTools for humans (AI assistance powered by Gemini) and DevTools for agents, both tech-led by Yang Guo and Mathias Bynens.

Debugging used to stop at page load, but the problems that actually hurt users, slow interactions, layout shifts, broken journeys, all happen after it. And as AI arrived, both people and agents needed DevTools to become an intelligent collaborator, not just an inspector.

DevTools moved from a load-time inspector to a tool for whole user journeys and Core Web Vitals, and then into AI-assisted debugging for humans and a surface that agents can drive, for a tool more than 40 million developers rely on.

I contributed to the developer-facing journey for Google's Gemini models, from the early multimodal work in Gemini through coding, partnering with Google DeepMind to make Gemini stronger at full-stack web development, and supporting launches like the Gemini CLI and Nano Banana image generation.

State-of-the-art AI only reaches developers if it fits naturally into how they actually build, in their editor, their terminal, their real workflow, not just a demo.

Helped put Gemini's capabilities into the hands of millions of developers through clear docs, hands-on demos, and reference apps, and helped define what agentic AI means for the people who build software.

As a Director at Google Cloud AI, I led developer experience and technical evangelism for Google's enterprise agent stack, the Gemini Enterprise Agent Platform, alongside Agent CLI, Agent Studio, the Agent Development Kit (ADK), Gemini, and Vertex AI.

Enterprises need more than a model endpoint to trust AI agents in production. They need a real platform, and a real developer experience, to build, deploy, evaluate, and govern them.

The organization hit its annual developer-growth goal, roughly 2x, in under six months.

An open collection of production-grade "skills" that make AI coding agents work with senior-engineer discipline, tests, review, verification, across the whole development lifecycle. Companion site at skills.addy.ie.

Coding agents generate fast but skip the discipline that makes software trustworthy. The skill is no longer writing the code; it's the constraints you wrap around the machine.

It became the most popular independent Agent Skills collection on GitHub, at around 90,000 stars, and counting.