Homecoming

Your taste didn’t move. Your address did.

Comment

Inspiration

Moving to a new city changes your address, but not your taste. Yet your favourite café, the band you always went to see, the film everyone in your circle loved: none of it comes with you. Generic "best of" lists don't help, because a five-star place can still feel like the wrong room. [Add one sentence about your own experience of moving or of someone you know.]

We asked a simple question: what in this city is the equivalent of the thing I loved back home? That is a question about taste, and it's exactly what Qloo's taste graph knows.

What it does

Homecoming translates your taste from the city you left to the city you moved to.

  1. You name two cities and a few favourites: a place, an artist or a film.
  2. Qloo finds what people with your taste love in the new city.
  3. Each match shows the traits it shares with your original, so you can judge the match yourself.
  4. Jev scores how close the match is, and the best three are kept.
  5. A short first-weeks plan turns the matches into things to actually do.

How we built it

  • Qloo resolves each favourite to an entity with /search, then /v2/insights returns places or artists and films that people with that taste love in the new city.
  • Jev 1.13 judges each pair. For every candidate it answers two yes/no probability questions, one about atmosphere and one about the role it plays in daily life. The fit score is their average:

$$\text{fit} = \frac{p_{\text{vibe}} + p_{\text{role}}}{2}$$

We fetch eight candidates per favourite, keep the three with the highest fit, and drop any below 0.3 unless that would leave nothing.

  • Nemotron writes the first-weeks plan from the matched places and shared tags only. If the model is unavailable, the plan is assembled directly from the matches, so it never invents a venue.
  • Frontend: Next.js, React and Tailwind, with a small design system and checked colour contrast. -Backend: FastAPI. Hosted on Vercel and Render.

The three AI pieces each do what they are best at: Qloo finds, Jev judges, the language model writes.

Challenges

  • Messy taste data. Early results carried tags like "Restroom" and age brackets, and repeated the same tag twice. We built a tag cleaner so the shared traits shown are actually about taste.
  • Weak matches. A hotel showed up as a match for a bakery. Fetching more candidates and letting Jev choose the best fixed most of it, and the labels "strong", "good" and "loose" keep the scores honest.
  • Jev only sees tags, not the venues themselves, so we treat its score as a grounded estimate of closeness, not a review.
  • Free-tier limits. The hosted language model returned a 403 error, which pushed us to build the no-model fallback plan.
  • Shipping. Deployment problems and a design that needed a full rebuild after feedback. Along the way we learned that Tailwind silently drops opacity classes on CSS-variable colours.

What we learned

Grounding matters more than model size. Once a language model is told to use only data from the taste graph, it stops guessing, and splitting the work between a retrieval layer, a fast scorer and a writer made each part easier to test and to fix. Building for failure was just as important: every AI call can fail, and the app still has to work.

What's next

Filtering results by venue type, more domains such as books and food, a "bring a friend" mode that compares a newcomer's taste with a local's, and saving a shortlist.

Built With

Share this project:

Updates

Submission history