Taste Concierge

Tell it what you love, get real places in a city that fit your taste, from Qloo. Then it tests its own answer against a model alone, scored by Qloo, and shows who wins.

Comment

Inspiration

Ask a chatbot where to eat and you get the same famous places every guide lists. We wanted an agent that starts from what you love, a film, a band, a book, and proves with numbers that it does better than a model alone.

What it does

You write one sentence, such as "I love Khruangbin and The Grand Budapest Hotel. Where should I eat in Lisbon?". A router model reads it with tool calling and fills a plan: the loves, the venue, the city. Qloo finds each love in its taste graph, reads the taste as ambience words (whimsical, rustic, bohemian), and ranks real places in the city by affinity, with the share of the fit from each love. A writer model writes the answer from the Qloo picks only; it may not add a place. "Run the test" asks the same model for places with no Qloo data, finds them in Qloo and scores both lists on the same loves. A held-out test then ranks again without one love and scores both lists on that love alone, because the Qloo list is ranked by the same affinity it is scored on. Follow-ups such as "Now record stores there." reuse the saved loves and city. The same agent is an MCP server with 4 tools, so other agents can use it.

Results on the live demo

Over 7 live requests on 2026-10-10, the Qloo list was less famous than the model-only list in 7 of 7, and Qloo won 10 of 14 held-out loves. It lost all 4 in Paris, where the model's well-known bookstores and museums fit Amélie and Patti Smith better, and the page says so. The raw run with every place and score is in the repo: submission/qloo/live-eval-2026-10-10.json.

How we built it

A Cloudflare Worker with Hono and D1 for per-browser taste memory. Qloo /search and /v2/insights (places with explainability, tags for ambience, filter.popularity.max for "skip the famous places", and filter.results.entities for scoring the model's places). NVIDIA Nemotron 3 Nano routes and Nemotron 3 Ultra writes, through Nebius Token Factory; until the Nebius key is set, the demo runs the same code on Workers AI llama-3.3-70b, and every answer names the model that wrote it. The OpenKrill travel MCP server adds official travel advice, and Tavily adds events this month. 58 offline tests replay recorded live calls, and all 58 pass.

Challenges we ran into

Qloo tags a place with every category it has, so a coworking space with a kitchen is also a restaurant; the agent uses the main category first and asks Qloo again for a deeper list when too few fit. A fair test was the hard part: a higher fit on the same loves is close to guaranteed, so we added the held-out test.

Accomplishments that we're proud of

The page reports honestly when the model-only list wins, and every number on it comes from Qloo, not from the model.

What we learned

Taste data changes the answer most where a model falls back on fame; where Qloo has little data, saying so beats guessing.

What's next for Taste Concierge

Run the live demo on Nemotron through Nebius, and add more venue kinds and cities.

Built With

  • cloudflare-d1
  • cloudflare-workers
  • hono
  • mcp
  • qloo
  • tavily
  • typescript
  • vitest
  • workers-ai
Share this project:

Updates

Submission history