TasteScout

Inspiration

Small businesses decide their menu, partnerships and social media by gut feeling. Big brands pay for consumer research; a specialty café in Mexico City can't. Qloo's taste graph knows how preferences connect across places, brands, music and media. We wanted to put that intelligence in the hands of a shop owner, in plain language, in under a minute.

What it does

You type your business (e.g. "specialty coffee shop"), your neighborhood (e.g. "Roma Norte, Mexico City") and, optionally, a few brands or artists your ideal customer loves (e.g. "Bad Bunny, Blue Bottle Coffee"). TasteScout's agent then:

1. Resolves your references into Qloo entities with /search.
2. Decides which taste dimensions to explore and queries Qloo Insights (/v2/insights) for places, brands, artists and podcasts, using your neighborhood and references as signals. Places are filtered to your actual area, so it surfaces real neighbors your customers already love.
3. Writes a grounded plan: who your audience is, 3 concrete things to add to your menu or catalog, specific local businesses, brands and artists to partner with, and the tone and topics for social media.

Compare two neighborhoods: deciding where to open is one of the most expensive choices a small business makes. In comparison mode, the agent runs the same Qloo queries for two locations (e.g. Roma Norte vs. Coyoacán), shows both taste profiles side by side, highlights what the two audiences share and where they differ, and recommends where to open and how to adapt the menu, partnerships and tone to each.

Each card is tagged "only in this zone" or "in both", so the real differences jump out, and places Qloo returns from a neighboring area are flagged (the agent is told too, so it doesn't present them as local).

From insight to action: every analysis ends with a 4-week social media calendar (12 posts, each inspired by a Qloo entity) and a one-page PDF with the plan and the calendar, ready to hand to the team.

The results are shown as visual cards (top places, brands, artists and podcasts with an affinity bar). Every recommendation cites the Qloo entities and affinity scores behind it, and anything that is the model's own interpretation is labeled as a suggestion. The app shows each agent step and every raw Qloo API call, so the owner can see exactly why the agent recommends what it does.

Example (real output): for a specialty coffee shop in Roma Norte, TasteScout found high affinity with artists like Blood Orange, Buscabulla and Jungle, wellness podcasts like On Purpose with Jay Shetty, brands like Nike Training Club and Herschel, and nearby venues like La Bohême and Salón Las Tertulias. It turned that into a functional wellness drinks line, a pastry partnership with La Bohême, "Morning Run & Coffee" events with local running clubs, and monthly in-store playlists.

How we built it

• Agent: an LLM with tool calling (OpenAI-compatible API, running on Google Gemini) and two tools wrapping Qloo's Taste AI API: entity search and insights. The model chooses which entity types and signals to query, iterates, and stops when it has enough data.
• Qloo integration: /search to resolve references into entity IDs; /v2/insights with signal.interests.entities, signal.location.query for taste signals and filter.location.query to keep places local.
• Grounding: the system prompt forces every claim to come from Qloo results and to mark interpretations as suggestions.
• Action plan: a second LLM call turns the plan into a 4-week posting calendar; the PDF is generated with fpdf2 and auto-scaled to fit one page.
• Front end and deploy: Python + Streamlit on Streamlit Community Cloud. The Qloo key stays server-side in Streamlit secrets.
• Resilience: automatic retries and a fallback model when the LLM is overloaded; a demo mode with sample data when no key is configured.

Challenges we ran into

• Choosing the right location parameter per entity type: for places we needed a hard filter to the neighborhood, while for brands, artists and podcasts the location works better as a taste signal.
• Keeping the LLM honest: the first version described the audience with traits Qloo never returned (e.g. job or commute habits). We tightened the prompt so the audience profile is built only from high-affinity entities and suggestions are labeled.
• Approximate location filtering: Qloo's place filter sometimes returns venues from neighboring areas, so we check each address against the requested zone, tag those cards and tell the agent.
• Free-tier LLM capacity: Gemini occasionally returned 503 (overloaded) in the cloud, so we added retries and a fallback model.

Accomplishments that we're proud of

• A working end-to-end agent in under 3 days, from idea to public deployment.
• Recommendations that name real neighboring businesses, which turns abstract "taste data" into partnerships an owner can act on tomorrow.
• A neighborhood comparison that helps answer one of the most expensive questions a small business faces: where to open.
• Full transparency: every recommendation can be traced back to a Qloo API call.

What we learned

• Cross-domain taste signals (music → food → brands → places) are far more actionable for a small business than demographics alone.
• Tool calling plus a strict grounding prompt is what separates a useful agent from a confident storyteller.
• Insights only matter when they become actions, which is why the plan ends in a calendar and a PDF.

What's next for TasteScout

• Compare more than two locations and rank them.
• Demographic signals (age, gender) to target a specific audience.
• A weekly "taste trends" digest for the owner.
• Follow-up chat to refine the plan ("how do I bring in more afternoon customers?").

Built With

• fpdf2
• google-gemini
• openai-sdk
• python
• qloo-api
• streamlit