v0.11 · Source-available
ScrollCollector
A Chromium browser extension that collects the posts in your feeds that match your own prompts, as you scroll. A small embedding model bundled with the extension decides each post on your device; the browser’s built-in model is asked only about the posts it is unsure of.
Widen the panel and the stream reflows into columns.
On deviceby default; cloud is opt-in
Social feeds and developer blogs are high-noise, high-density environments. Curating insights manually is slow, and sending every scrolled post to a cloud LLM compromises user privacy while adding latency and API costs.
The problem
ScrollCollector decides inside the browser.
It scrubs emails, card numbers and phone numbers from the text first, and post text leaves the device only if the user turns on a cloud provider.
What I built
- A decision cascade. A q8 sentence-embedding model (all-MiniLM-L6-v2 by default, bge-small selectable) runs with Transformers.js on WebAssembly in an offscreen document and decides every post. The weights ship inside the extension; nothing downloads at runtime.
- The built-in model only when unsure. Low-confidence posts go to the browser’s Prompt API; a rule-based classifier answers whatever is left, offline.
- Prompt lenses. Several prompts at once, each with its own collect switch, and ranked views that show which prompt a post matched, best match first.
- Opt-in cloud. TypeSafe Jev and OpenAI are off by default, each with its own opt-in and key, and only uncertain posts go to one provider at a time.
- Manifest V3 extension. Scroll-optimized content scripts find posts by stable React data attributes rather than obfuscated classes, on LinkedIn, X, YouTube and more.
- Virtualized side panel and HTML diary export. Off-screen cards become height-matched spacers, and the collection exports as a standalone, searchable HTML diary.
- A Go installer. It detects Chrome, Edge and Brave and walks through loading the extension.
What I learned
Handling dynamic social feeds means dealing with frequent class name obfuscation (such as on LinkedIn or X/Twitter).
Relying on traditional CSS selectors fails immediately. Switching to stable React internally-exposed attributes (like data-urn values) proved to be the only reliable way to maintain long-term scraper resilience.
Additionally, updating complex state in a narrow browser sidepanel during fast scrolling causes DOM thrashing. List windowing and IntersectionObservers are not just optimizations — they are architectural requirements to keep chromium extension processes from crashing.
Technical highlights
- 05842
842 unit tests on Vitest and 42 end-to-end tests on Playwright, all passing at the v0.11 release.
- 01
Bundled sentence-embedding model run with Transformers.js on WebAssembly in an offscreen document: no model download at runtime.
- 02
Escalation to the browser Prompt API (Gemini Nano, Phi-4-mini) only for low-confidence posts, with a rule-based fallback.
- 03
Manifest V3 Chromium extension with background service-worker lifecycle control.
- 04
IntersectionObserver-based list virtualization to prevent DOM memory overflow.
By the numbers.
- 3 tiers
- embeddings, built-in model, rules
- v0.11
- prompt lenses and ranked views