Truce

Truce_full logo
Truce_text

Inspiration

Bridging political or tribal division rarely works through debate. It works through shared fringe passions — a song, a film, a place — that let people meet as humans instead of avatars. We also noticed recommendation engines and LLMs default to either echo chambers or bland “meet in the middle” advice. Qloo’s taste graph promised something different: entity-backed cultural affinity across music, film, brands, and destinations. The question was whether that graph could power an agentic product for a real social job — a dinner, a community meet-up, a host who needs a bridge card tonight.

What it does

Truce is a working web app where you fill two taste tables (media, movies, artists, food, holidays). It calls Qloo Insights and returns:

A bridge list of shared entities with both-side affinity scores
Nostalgic prompts (“Who showed you this artist first?”) instead of critique
Dinner-table games (Blind Pick, Watch Swap)
A static Left only / Shared / Right only Venn of real entity names (no fake percentages)
An optional Truce Deck — flashcards with Love / Hard Pass; when both sides Love, “Common ground locked”

The bridge is computed primarily from media-diet signals (the path that actually differentiates in Qloo). Artists and movies on the cards are profile context — feeding opposing artists as signals pulls the graphs apart.

How we built it

Node + Express server; static front end (no SPA framework)
Qloo Search to resolve names → Insights (signal.interests.entities) for artists, movies, brands, destinations
Server-side API key only; serial Qloo queue + 429 backoff (hackathon keys rate-limit hard under parallel bursts)
Demo presets for Blue/Red media diets; editable chips with +
Hosted intent: Railway + public GitHub (MIT)

Probe scripts and link document what we learned against hackathon.api.qloo.com before locking the product shape.

Challenges we ran into

political_preferences audiences (Progressive vs Conservative) returned identical Insights lists and a flat affinity (~0.765). Cultural media brands as signals differentiated immediately.
Mixing opposing artists into Insights signals erased the Americana overlap we’d found with media alone.
HTTP 429 from parallel Insights calls looked like “no shared bridges” until we logged status codes and serialized requests.
Place/filter.location city strings failed WKT parsing; we scoped destinations/brands instead of unreliable place geo for the demo.
Shipping a complete host experience (agenda, Venn, deck, games) in a very short window without turning into a generic chat wrapper.

Accomplishments that we're proud of

A demo judges can use end-to-end, not a slide deck
Product honesty about how Qloo is used (media diets → bridge; not magical audience URNs)
Human layer: memory prompts + games + optional deck celebration
A concrete FEEDBACK.md for Qloo on audiences, signals, rate limits, and overlap DX

What we learned

Qloo’s strength here is cross-domain affinity from cultural entities, not “political audience in → consensus out” on this hackathon surface.
Empty result arrays after a burst often mean rate limit, not empty taste.
Agents with taste still need a table-shaped UX — affinity scores alone feel clinical; autobiography unlocks the room.

What's next for Truce

Hosted multi-user rooms (two phones, one deck)
Deeper destination/place bridging once geo params are rock-solid on hackathon
Optional Spotify / JustWatch deep links on bridge cards
Community / civic host packs (libraries, dinner series) with preset table templates
Share learnings with Qloo on a first-class overlap helper and clearer 429/audience docs