RSS Tinder

RSS feeds don’t seem to be en vogue anymore, apparently. Everyone is on algorithmic timelines now. But I still love the control of RSS; except for the part where I have 1,000+ unread items and a mild panic attack every time I open my reader.

So I built RSS Tinder. RSS readers are great at aggregating content but terrible at filtering it. What I wanted is known from dating Apps: Instead of a list, give me cards. Swipe right to like, left to dislike. Then, train a model to do the heavy lifting for me.

So I built the frontend with vanilla JavaScript because I wanted it to be lightweight and fast. No React, no build steps, just raw DOM manipulation and CSS transitions. It’s actually quite satisfying. You drag a card, it follows your mouse (or finger), and snaps away with a satisfying thwack. Behind the scenes, the app tracks everything. Not just votes (Like/Neutral/Dislike), but engagement time and link opens. If I swipe right but don’t open the link, that’s a signal. If I spend 0.5 seconds on a card before swiping left, that’s a signal.

I started with XGBoost and some engineered features (title length, word count, time of day). It was decent—about 66% ROC-AUC. It learned that I generally dislike short, clickbaity titles. But it didn’t understand context.

I upgraded the pipeline to use MPNet (specifically all-mpnet-base-v2 from Hugging Face) to generate 768-dimensional embeddings for every article title and description. I fed these, along with my engineered features, into a Hybrid Random Forest model.

The result? 73.5% ROC-AUC. The model now understands that I like “Machine Learning” and “System Design” but gets bored by “Crypto” and generic “Startup Advice.”

 1
 2
 3
 4
 5
 6
 7
 8
 9
10

def predict_preference(article):
    # 1. Generate semantic embeddings
    embeddings = mpnet.encode(article.text)
    
    # 2. Extract behavioral features (time, author, feed)
    features = extract_features(article)
    
    # 3. Predict with Hybrid RF
    score = model.predict_proba(np.hstack([embeddings, features]))
    return score

The only problem with heavy ML models is latency. Scoring an article with MPNet took about a second. In a swipe interface, a one-second lag feels like an eternity.

I solved this with a Preload Queue. The backend is frantically scoring and fetching the next 5 articles in the background.

1
2
3
4

if (cardQueue.length < 5) {
    fetch('/api/posts/batch?limit=5')
        .then(posts => addToQueue(posts));
}

Now, the UI is buttery smooth. It uses Thompson Sampling (80% exploit, 20% explore) to decide what to show you next. Most of the time it shows what it thinks you like, but occasionally it throws in a wildcard just to see if your tastes have changed.

Backend: Python & Flask (Simple, effective). Database: SQLite (Because Postgres is overkill for my reading habits). ML: Scikit-learn, XGBoost, Sentence-Transformers (MPNet). Frontend: Vanilla JS, CSS Variables.

If there is some interest I will make the code public so you can train and run your own version