OmniFeed

Your algorithm, across platforms.

Browser extension that intercepts your social media feeds, scores content against your personal ML preference model, and reranks in real-time. The model learns from your behavior on every platform you use.

Status: Archived proof-of-concept. Works on Twitter/X but is not actively maintained. The architecture is designed to support any platform — adding a new one is just writing an adapter.

Quickstart

1. Start the server

cd server
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
# First run downloads ~90MB sentence-transformer model
uvicorn app.main:app --reload --port 8000

Verify: http://localhost:8000/health → {"status": "ok"}

2. Build the extension

cd extension
npm install
npm run build

3. Load in Chrome

1. Go to chrome://extensions
2. Enable "Developer mode" (top right)
3. Click "Load unpacked"
4. Select the extension/dist/ directory

4. Test it

1. Navigate to https://x.com
2. Open DevTools console (F12)
3. You should see:

[OmniFeed] API hooks installed
[OmniFeed] Bridge initialized for twitter
[OmniFeed] Feed observer attached

4. Scroll your feed — as tweets load:

[OmniFeed] Extracted 25 items from twitter
[OmniFeed] Scored 25 items in 47ms
[OmniFeed] Reranked 22 tweets

5. Monitor

• Popup: Click the OmniFeed extension icon → shows server status + interaction count
• Stats: http://localhost:8000/api/v1/user/dev_user/stats
• Training log: http://localhost:8000/api/v1/debug/training-log
• Cache: http://localhost:8000/api/v1/debug/embedding-cache

Architecture

x.com (or any supported platform)
  │
  │ Twitter GraphQL API response lands in browser
  │
  ▼
loader.ts (MAIN world)
  │ Patches XHR/fetch, captures timeline responses
  │ Fires CustomEvent with raw JSON
  │
  ▼
bridge.ts (Isolated world)
  │ Receives event, runs Twitter adapter
  │ Adapter extracts tweets → CanonicalItem[]
  │ Processor converts → WireItem[] (with author hash, bucketed metrics)
  │ Sends to server for scoring
  │
  ▼
Server (FastAPI, localhost:8000)
  │ Embeds text via sentence-transformers (384-dim)
  │ Scores against user's preference vectors:
  │   core + temporal(time-of-day) + platform + author_affinity + type + freshness
  │ Returns ranked scores
  │
  ▼
bridge.ts
  │ Receives scores, calls adapter.reorderDom()
  │ Tweets physically reorder in the DOM
  │
  ▼
Tracker (IntersectionObserver + click delegation)
  │ Monitors dwell time, likes, replies, shares
  │ Every 15s, flushes interaction signals to server
  │
  ▼
Server: PreferenceEngine.update()
  │ Computes reward from signals (relative to user's baseline)
  │ Nudges preference vectors toward engaged content
  │ Logs full training record for future model improvements

Adding a New Platform

1. Create extension/src/content/adapters/{platform}.ts
2. Implement the PlatformAdapter interface (see shared/types.ts)
3. Map platform-specific fields → CanonicalItem (same shape for everyone)
4. Add URL patterns to loader.ts
5. Add host permissions to manifest.json
6. Register adapter in bridge.ts

Everything downstream (processor, server, ML) works unchanged.

Project Structure

omnifeed/
├── extension/
│   ├── manifest.json                    # Chrome Manifest V3
│   ├── package.json                     # Vite + CRXJS
│   ├── vite.config.ts
│   ├── tsconfig.json
│   └── src/
│       ├── content/
│       │   ├── loader.ts                # MAIN world: XHR/fetch interception
│       │   ├── bridge.ts                # Orchestration: adapter → server → rerank
│       │   ├── adapters/
│       │   │   └── twitter.ts           # Twitter/X GraphQL → CanonicalItem
│       │   └── core/
│       │       ├── processor.ts         # CanonicalItem → WireItem
│       │       └── tracker.ts           # Dwell + interaction tracking
│       ├── shared/
│       │   └── types.ts                 # All types + PlatformAdapter interface
│       ├── popup/
│       │   └── index.html               # Extension popup
│       └── background/
│           └── index.ts                 # Service worker (minimal)
└── server/
    ├── requirements.txt
    └── app/
        └── main.py                      # FastAPI + Embedder + PreferenceEngine

License

MIT