TasteSphere

A multi-domain personal assistant that recommends where to travel, what to buy, and how to live — based on your unique tastes — using Qloo + Claude.

Comment

Inspiration

We were inspired by the idea that cultural discovery can be more personal, meaningful, and focused. While many platforms offer broad recommendations, we wanted to create a system that allows users to explore deeply within one interest type at a time — whether it’s movies, places, video games, or beyond — powered by real cultural data and natural language understanding.

What it does

TasteSphere is a personalized cultural discovery tool. The user selects a single interest type (e.g., movie, place, or podcast) and then enters a free-form prompt describing their tastes, feelings, or preferences. TasteSphere returns curated results from that specific domain, grounded in real data.

Example:

“I’m looking for restaurants in NYC that offer brunch, vegan options, and a wheelchair-accessible entrance.”

TasteSphere parses the intent, queries Qloo to find relevant content in the selected domain, and generates a rich, readable response — complete with summaries, visuals, and more.

How we built it

  • Backend: Python + Django
  • Frontend: React + Tailwind CSS
  • AI Stack:
    • Claude 4 Sonnet (via AWS Bedrock) is used in two places:
    • To parse the user's natural-language prompt and extract parameters like:
      • signal.interests.entities
      • filter.location.query
      • filter.tags
    • To generate the final user-facing response, based only on the selected interest type and Qloo insights.
  • Cultural Data: Qloo API is used in two steps:
    1. Called with extracted keywords to fetch relevant entity and tag IDs
    2. Called again with those IDs + interest type to get a curated list of recommendations
  • Output Generation: Claude 4 Sonnet takes both the Qloo data and original user intent to generate the final, grounded, helpful recommendation.

Challenges we ran into

  • Building a multi-step AI pipeline where Claude is used twice in different roles (parsing and generation)
  • Mapping vague natural language into structured query parameters acceptable by Qloo
  • Handling Qloo edge cases where data was missing or sparse
  • Making sure the final AI response remains grounded in Qloo’s data, not general hallucination

Accomplishments that we're proud of

  • Integrated Claude 4 Sonnet with Qloo in a seamless, structured prompt flow
  • Developed a robust backend using Django to orchestrate multi-stage data flow
  • Created a system that makes focused discovery more creative and intelligent
  • Built a clean and extendable frontend that shows insights in a compelling way

What we learned

  • The power of using language models to bridge user input and structured APIs
  • How to split the roles of a model between understanding and generation
  • Real-world API integration with cultural data sources like Qloo
  • The value of separating AI reasoning and data retrieval into modular components

What's next for TasteSphere

  • Explore voice and image-based input
  • Enable daily AI-generated discovery emails
  • Add Google Authentication and user profile saving to keep track of favorites

Built With

Share this project:

Updates