Taste Concierge 🎬🍽️🎵🎁

Three answers in, one culturally coherent evening out — powered by the Qloo taste graph.

Most agents answer well. Taste Concierge knows what people like: it maps a 3-question interview to Qloo's taste-graph entities (250M+), then chains insights across dining, film, music and retail to compose an evening that fits that person — not the charts.

Why this isn't "an LLM with extra steps"

An LLM alone recommends popular things. Taste Concierge recommends that person's things:

1. Interview → entities. Your answers ("Studio Ghibli, Radiohead") are resolved to real taste-graph entities via Qloo /search.
2. Cross-domain chaining. Each slot's /v2/insights call is seeded by your entities — and the playlist is additionally seeded by the chosen film, so the music follows the night's centerpiece. That's cultural coherence, not four independent top-10s.
3. Provenance on every pick. Each card shows why it was chosen (which affinities connected it). No black-box recommendations.
4. A live feedback loop. Dismiss any pick — the plan recomposes instantly with the rejection recorded in your entity profile. Negative signals are first-class.

Remove Qloo and the product collapses: without the taste graph there is no entity affinity, no cross-domain coherence, and no grounding — just generic popular answers.

Run it

# 1. Get a hackathon API key: see the Devpost overview page / docs.qloo.com
# 2. Configure and start
cp .env.example .env        # put your key in QLOO_API_KEY
npx vercel dev              # local dev server

Deploy:

npx vercel --prod           # set QLOO_API_KEY in the Vercel dashboard

Architecture

index.html          the interview UI + plan cards (zero-framework, fast)
api/plan.ts         serverless endpoint: the agent loop
src/lib/qloo.ts     Qloo client (X-Api-Key, GET-only, hackathon base URL)
src/lib/agent.ts    the pipeline: mapAnswersToEntities → composeEvening
                    → recomposeWithRejection (the negative-feedback loop)

The agent pipeline is deliberately small and inspectable — every Qloo call and every affinity decision can be traced in the code.

Hackathon

Built for the Qloo Agentic Hackathon. License: MIT.