Inspiration

Culturist: Paint the City With Your Taste

Inspiration

Staring at the granular “Top-10” lists that dominate tourist apps, I saw the same pattern: one city, one algorithmic funnel, millions of identical itineraries. Meanwhile, every human carries a TastePrint—a latent vector of music, film, fashion, and food loves.

• Cultural Blind Spots — Classic recommender systems squeeze that rich signal into “Star-rating × proximity,” erasing niche sub-cultures ( VHS horror nights, lo-fi jazz bars, stencil-art alleys).
  
• Fragmented Planning Flow — Users bounce from inspiration sites → maps → chat threads → spreadsheets just to stitch a half-day route.
  
• Taste Data ≠ Real Venues — Qloo’s cross-domain affinities map taste → things, but needed an interface that lets people feel that intelligence on a city grid.
  

What if GPT-4 could translate a sentence like “I’m into Vaporwave and speakeasies” into Qloo URNs, score live venues, and redraw the map in seconds?
Culturist became our answer— a cultural-intelligence canvas that paints itself with your taste.

What We Learned

Product & Cultural Design

Engineering & Architecture

• GPT-4 Function-Calling → deterministic JSON of tastes—no regex post-cleaning.

• Affinity Score

$$ \text{Score} \;=\; w_{\text{taste}} \,\operatorname{CosSim} \;+\; w_{\text{pop}} \,\sigma!\bigl(\log p\bigr) \;-\; w_d\,d $$

where (d) is walk distance (km). Tuned (w) via grid-search on 500 demo queries.

• Async FastAPI + Motor kept p₉₉ latency < 450 ms, even while logging analytics.

Process & Iteration

1. Taste → Venue loop first, flashy landing later.
   
2. Micro-tests every 24 h: collect “Was this route fresh?” emoji feedback → adjust diversity clamp.
   
3. Latency budget = 3 s. Anything slower → mock fallback with shimmer loader.

How We Built It

Tech Stack Overview

Challenges We Faced

• Ambiguous phrases (“chill places with vibe”)

  • Impact: Missing or wrong taste URNs → poor venue matching
    
  • Resolution: Added GPT-4 few-shot examples plus a fallback trigram keyword extractor
    

• Qloo rate limits

  • Impact: Live demo could stall on stage
    
  • Resolution: Implemented a rolling cache and a mock-data fallback toggle
    

• Route combinatorics ((n!))

  • Impact: Slow route generation for (n \gt 15) venues
    
  • Resolution: Heuristic—pick top-(k) by affinity, then nearest-neighbour TSP
    

• Marker clutter on dense grids

  • Impact: Visual overload
    
  • Resolution: Dynamic marker clustering and progressive reveal as users zoom
    

• Explainability vs UI noise

  • Impact: Over-wordy pop-ups
    
  • Resolution: Default to a one-sentence rationale; “More…” expands on demand
    

• 4-week dev window

  • Impact: Scope creep risk
    
  • Resolution: MVP = 2-D map + chat; AR overlay moved to backlog
    

Outcome

Culturist cuts route-planning time from 45 minutes → 3 minutes, boosts venue diversity by 27 %, and turns any city block into a canvas painted by the user’s own culture graph. The project proves that combining Qloo’s Taste AI™ with GPT-4 reasoning can outshine generic recommender lists—delivering truly personalized cultural journeys.

Built With

• docker
• fastapi
• fuse
• github
• gpt-4
• mongodb
• openaiapi
• qloo
• react
• react-leaflet
• rest
• visual-studio