Sift

About

Sift runs EmbeddingGemma-300M (q4) directly in your browser to score feed items against your interests and fade low-relevance posts. All inference happens locally — no data ever leaves your machine.

SUPPORTED SITES

- Hacker News
- Reddit
- X (Twitter)

HOW IT WORKS

1. Pick scoring categories — 25 built-in across tech, world, and lifestyle
2. Browse normally — Sift embeds every title and scores it against your categories
3. Low-relevance items fade, high-relevance items stay vivid
4. Category pills show which topics match each item

FEATURES

- WebGPU acceleration with WASM fallback
- Score inspector — click "?" to see why an item scored the way it did
- Per-site toggles and sensitivity slider
- Auto-detected category pills in the popup and feed
- Light/dark mode (follows system)

TASTE PROFILE
After labeling 10+ items with thumbs up/down, Sift builds a contrastive taste profile showing your top interests ranked by affinity, with an interactive radar chart.

TRAINING LOOP

- Label items with thumbs up/down as you browse
- Curate labels in the Label Manager — edit, flip polarity, reassign categories
- Export as CSV training triplets
- Fine-tune the model with the included Python pipeline or free Colab notebook
- Load your fine-tuned model back into the extension

PRIVACY-FIRST

- No backend server, no analytics, no telemetry
- All inference runs locally in your browser
- Labels and settings stay in local storage
- Model weights are the only network download (from HuggingFace Hub)

Open source (Apache-2.0): <https://github.com/shreyaskarnik/Sift>