Lume geliştiren: Digitalhen

What it does
Lume is a "what to read next" engine that runs entirely in your browser. As you read articles around the web, Lume notices when you've dwelled on a page (5+ seconds, at least one scroll) and embeds it locally using a small sentence-transformer model (Xenova/all-MiniLM-L6-v2, ~25 MB, int8-quantized).

A live US news feed refreshes hourly behind the scenes — general, technology, business, and science top headlines from major outlets. Lume ranks those stories against what you've been reading recently and surfaces them in two places:

• An in-page popup that slides in bottom-right after you've dwelled: "Read next: …"
• A Discovery page (toolbar icon) with a grid of 10 stories matched to your reading, diversified so you don't see ten variations of the same story, each annotated "Because you read X"

What's local vs. remote
- Local: page text capture, embeddings, similarity matching, your reading history, the "muted sites" list. None of this ever leaves your device.
- Remote: a one-time ~25 MB model download from huggingface.co (cached forever); periodic news-feed fetch from a Cloudflare Worker we operate (the news API key stays server-side; no user data is sent in the request).

No analytics. No telemetry. No accounts.
Source code: https://github.com/digitalhen/story-discovery-localai