ProductivityBoost

This Chrome extension case study covers a developer-tools product built for a Y Combinator batch company. The problem: developers were losing an average of 23 minutes regaining focus after each interruption, and existing focus tools were either too aggressive or effectively invisible.

The solution was an ambient AI focus assistant: an on-device TensorFlow.js model classifying distraction patterns with no data leaving the browser, smart tab hibernation cutting memory use by 60%, and contextual focus blocking tied to the user's calendar and current project.

Within 90 days of launch the extension passed 50,000 active users, was featured in the Chrome Web Store's Editor's Choice, and held a 4.8-star store rating, with user retention rising from 12% to 68% and background CPU overhead kept under 1%.

The build ran over twelve weeks with a team of two engineers. The first two weeks covered user interviews with fifty developers, a cognitive-load analysis and a teardown of the existing focus tools on the market. Weeks three to five built the machine learning pipeline: labelling a training dataset, training the TensorFlow.js model, and quantising it to run inside a browser without a noticeable performance cost. The extension core, its background service worker and Chrome API integrations, took a further four weeks, followed by a beta with 500 developers and NPS surveys before the public Chrome Web Store submission and a Product Hunt launch in week twelve.

The stack was picked specifically for a zero-permission, on-device model: TypeScript and React for the interface, TensorFlow.js for on-device inference so no user data ever leaves the browser, Chrome's own extension APIs for tab and window management, and a local SQLite database compiled to WebAssembly for the stats dashboard, with Webpack 5 handling the production build.

A few details from launch are worth calling out specifically, since they are what actually got measured rather than assumed. The extension held under 1% CPU overhead while running in the background, memory usage dropped from roughly 800MB to 320MB through the tab-hibernation feature, users settled into an average of 4.2 focus sessions a day once the assistant was active, and the entire distraction-classification model runs on-device with a zero-permission design, so no browsing data is ever sent to a server.

None of this was built as a generic productivity tool. Every feature maps to one specific pain point from the original developer interviews, and the team judged the project by whether those exact numbers moved, not by how sophisticated the underlying model looked on its own.

See Same-origin policy (web.dev) for the security boundary an on-device, zero-permission model like this one has to respect. Our web development and Facebook ads services cover the same kind of work described here.