Inspiration
City recommendations usually split into two extremes: a chatbot that invents vibes, or a map that filters by category and distance. Neither answers the question we actually ask when traveling or exploring a new neighborhood — "If I love Miles Davis / Blade Runner / this book, where should I look nearby?"
That gap inspired CityPlay AI. Cultural taste is already a rich signal; cities are already full of places. What's missing is an agent that can connect them with receipts — real entities, real associations, and places you can verify on a map.
What it does
CityPlay AI turns a favorite musician, film, book, or concept into grounded nearby place candidates.
- Parse a natural-language request (or take a direct cultural name) into a short exploration plan.
- Resolve cultural anchors to real Qloo entities (and tags when names don't map cleanly).
- Query cross-domain taste associations within a chosen venue direction and area.
- Check category and coordinate constraints.
- Present a few candidates on an interactive map — each with its source and why it was kept or dropped.
The core experience is Parallel Lives: one input, three versions under the same city center, radius, categories, and hard constraints.
- Familiar Me — places matched to the interests you confirmed.
- Adjacent Me — the agent asks Qloo for an associated artist/cultural bridge, re-validates that entity, then searches places from the bridge, showing interest → bridge → place.
- Alternative Me — you supply a new starting point; the same resolution and checks apply.
Where versions genuinely overlap, we show the overlap. We never fabricate differences to make personalization look stronger than it is.
How we built it
- Qloo Culture API for
/search,/v2/tags, and/v2/insights(with explainability) — every entity ID, tag, coordinate, and candidate comes from Qloo. - DeepSeek (plus selectable OpenAI-compatible presets) only structures natural-language intent. A direct name-entry path keeps the core Qloo loop usable without the model.
- Python standard-library HTTP server, SQLite cultural library, threaded jobs, and bounded Qloo pacing.
- Vanilla JS + Leaflet for the map UI — no heavy frontend framework.
- An optional same-category baseline keeps area, venue filters, and checks fixed while removing cultural signals, so judges can see how taste changes discovery.
Live demo: https://cityplay-ai.onrender.com/
Video: https://www.youtube.com/watch?v=AjQtP9XN0Rw
Source (MIT): https://github.com/millerice/cityplay-ai
The public demo is bring-your-own-key: open ⚙ API keys, pick a built-in model, paste Qloo + model keys. Keys stay in sessionStorage for this browser session (survive refresh; clear when the browser closes) and are never written to durable server storage.
What we learned
- Taste is not a prompt. Removing Qloo removes the candidate source. An LLM can phrase a plan; it cannot invent venue IDs or recommendation records we would trust.
- Visibility beats slogans. Showing resolved anchors, bridge hops, keep/drop reasons, and a same-category ablation made the product feel "Qloo-powered" far more than calling the workflow an agent.
- Honest overlap is a feature. Parallel versions often share places. Forcing uniqueness would have been fake personalization.
- Coverage has geography. New York is our primary verified demo area; Shanghai remains experimental, with weaker coverage in some niche venue categories.
- Empty results are data, not UI failure. When discovery returns nothing after a real Qloo path, saying so is better than widening constraints silently.
Challenges we faced
- Making cultural intelligence observable. Early builds looked like "NL → place search." We had to surface entity resolution, adjacent bridges, evidence cards, and the ablation baseline so the Qloo difference was undeniable.
- Entity ambiguity and tags. Not every name maps cleanly. Fallback to semantic tags and re-validation of bridge IDs became part of the agent loop, not an afterthought.
- Hard constraints vs. empty sets. Budget, open-now, exclusions, and radius must not be silently relaxed. Bounded re-plans and clear empty states took more design than "just return something."
- BYOK public demo without baking secrets. Contest hosting can't ship private keys; session-scoped client keys + server-side model base URLs was the privacy posture that still lets judges run end-to-end.
- Keeping the agent honest. Feedback ("not for me", "too far") re-ranks or re-searches this session only. We refused to invent broader dislikes, live hours, ticket prices, or route times we had not verified.
What's next
Cross-session taste memory, itinerary/route ordering, and richer per-candidate explainability as Qloo insights expose more signal — always with the same rule: cultural intelligence stays grounded, and unverified claims stay out of the cards.
Built With
- css
- deepseek
- html
- javascript
- leaflet.js
- openai
- openstreetmap
- python
- qloo
- render
- sqlite