VibePilot

Agents today can reason, plan, and book — but they are culturally blind. Ask one to plan a night around your taste and you get five unrelated top-ten lists. VibePilot is an autopilot for your taste: give it one line about what you love, and it builds an evening where every pick rhymes.

How it works. Your words resolve into seeds inside Qloo's taste graph — 250M+ entities across film, music, food, fashion, books, and places. A coherence engine scores every candidate: 60% seed affinity, 30% coherence with picks already chosen, 10% popularity. Below the floor? Dropped — never padded. Every surviving pick explains itself with its real shared taste tags.

Three real runs (shown in the demo video):

Date night in Paris, seeded with Wes Anderson and Radiohead: Jacquemus, Montmartre at dusk, Café Lumen, The Grand Budapest Hotel (coherence 0.46), Sufjan Stevens, The Elegance of the Hedgehog.
Solo recharge, seeded with Perfect Days and Brian Eno: a Kyoto tea house, The Little Book of Hygge, the Philosopher's Path, Paterson, Nils Frahm.
Food crawl, seeded with Everything Everywhere All at Once and Kendrick Lamar: Seoul Smoke BBQ, Smash & Sons, Shinjuku, Taco Truck El Sol, a Charli XCX playlist — maximalist, but coherent, with no repeated picks.

Qloo integration. VibePilot calls Qloo's search API to resolve seeds into entity URNs, then Qloo's insights API (filter.type across restaurants, movies, music artists, books, fashion brands, destinations; signal.interests.entities = seed URNs) for cross-domain recommendations, and uses tag overlap from Qloo's taxonomy for the coherence scoring. The live client (qloo_client.py) is test-covered with mocked transport; the QLOO_API_KEY env var switches it on. The demo video runs on the built-in taste graph (58 entities, 3 taste clusters) so anyone can run it with no key — the video states this honestly on screen. An API key has been requested through the hackathon's official form and live mode activates the moment it arrives.

Built and proven. Pure Python 3, standard library only, fully offline, $0 to run. 24 automated tests, all passing. A JSON output mode makes VibePilot a 'taste subroutine' other agents can call.

Honest limitations. The bundled graph is illustrative (live mode removes the limit); Qloo search results are tag-light versus insights, so pairwise coherence compensates; city is a soft preference, not a hard geo-filter.