Inspiration

Most recommendation systems are pattern matchers. "You read article A, here's similar article B." They track clicks, not culture. They see behavior, not taste. We wanted to build something that doesn't just recommend content—it profiles readers through the cultural entities they engage with, predicts their taste across categories, and gives publishers the intelligence to monetize that knowledge.

The challenge prompt asked: "Build the thing that only works because it knows what people love." That's exactly what we did. Without Qloo's cultural intelligence, this would just be collaborative filtering. With it, it's a propaganda-grade profiling engine.

What it does

Blogaganda is a news platform with two faces:

For readers: Personalized article recommendations that feel like mind-reading. Read about Victoria's Secret, get recommendations for Chanel and Calvin Klein articles.

*For publishers(admins) (the real product): * Deep cultural intelligence showing:

  • Demographics: Age groups, gender skew predictions
  • Brand affinities: Which brands/people/places your readers love
  • Profile evolution: Watch taste drift over time (affinity score changes)
  • B2B revenue intelligence: Content strategy, virality predictions, ad targeting values, CPM estimates
  • Geo-cultural fit: Regional affinity scores (NYC, Paris, London, LA, Milan, Tokyo)

The flow:

  1. User reads article → System extracts entities (people, brands, places, movies, TV shows)
  2. Entities enriched via Qloo → Cross-category taste profile builds
  3. Recommendations scored by affinity (0.93+ threshold)
  4. If < 3 matches, dynamic fetching pulls fresh articles from NewsAPI based on predicted entities
  5. Admin sees everything: who readers are, what they'll engage with, how to monetize

How we built it

Stack:

  • Frontend: React + Vite (Vercel)
  • Backend: Express (Vercel serverless)
  • Database: Supabase (PostgreSQL)
  • NLP: OpenAI GPT-4 for entity extraction
  • Cultural Intelligence: Qloo API (Search, Insights v2, Demographics)
  • News: NewsAPI + NewsData.io

Architecture:

  1. Entity Extraction: Articles fetched → OpenAI extracts named entities → Qloo Search API matches to entity IDs → Only articles with Qloo entities saved

  2. Taste Profiling: User reads → entities added to profile → Qloo Insights API returns 15 cross-category predictions (3 per category) → Demographics API predicts age/gender → Profile evolution tracked

  3. Recommendations: Score articles by entity overlap with Qloo affinities → Dynamic fetching if < 3 matches (searches NewsAPI for predicted entities, enriches real-time)

  4. B2B Dashboard: Aggregates profiles across users, calculates virality scores, brand targeting opportunities, CPM values

Challenges we ran into

Cold start: Profiles initially required 3+ articles for recommendations. Killed engagement. Solution: Reduced to 1 article minimum + dynamic fetching.

Entity quality: 40% of articles had no recognizable entities. Solution: Improved OpenAI prompts, added title/description context, filtered zero-entity articles during seeding.

Scoring threshold: 0.3 was too loose, 0.95 too strict. Solution: Tuned to 0.93 after testing. Dynamic fetch acts as safety net.

Performance: Sequential API calls were slow (20+ seconds). Solution: Parallel batches of 5 for entity enrichment, reduced dynamic fetch to 1 article per entity.

Qloo quirks: Deprecated endpoints, URN format requirements, take vs limit parameters. Read the docs carefully.

What we learned

Technical:

  • Cross-category taste intelligence >>> single-domain recommendations
  • OpenAI + Qloo is a killer combo: LLM extracts, Qloo validates + enriches
  • Real-time enrichment is viable with caching + parallel processing

Product:

  • Users care about what they get (personalized content)
  • Publishers care about why it works (intelligence they can monetize)
  • The "propaganda" framing is uncomfortable but honest—every platform does this, we just made it visible

Qloo-specific:

  • The API is the differentiator. Without it, this is basic collaborative filtering
  • Demographics + Insights v2 unlocks real cultural profiling
  • 250M+ entity graph makes cultural resolution tractable

Built With

Share this project:

Updates

Submission history