TasteMenu
Know what your diners love. TasteMenu is a taste strategist for independent restaurants. An owner describes their place in one sentence. TasteMenu reads what people who eat at places like theirs also love (films, shows, music, brands, nearby places) using Qloo's taste graph, then turns that into this month's plan and the ready-to-send material to act on it.
Live demo: https://tastemenu-1037019152152.asia-south1.run.app (no login; tap a sample restaurant to start)
Built for the Qloo Agentic Hackathon.
The problem
Big chains pay agencies to understand their customers. A neighbourhood restaurant (a darshini in Bengaluru, a trattoria in the West Village) has no such help. Generic marketing advice ("play soft music", "run a weekend special") is the same for every restaurant, because it knows nothing about the people who actually eat there.
Qloo knows what audiences love across domains. TasteMenu uses that to answer the owner's real questions: what should I put on the menu, what should I play, what event should I run, who should I partner with, and which of my dishes should I push or drop?
What it does
| Feature | What the owner gets | Qloo tools behind it |
|---|
| Your crowd | The places most like yours in your city, and the films, shows, music artists and brands their audience over-indexes on, with Qloo affinity scores and an age/gender lean | find_tags, recommend, audience_demographics |
| This month's plan | Four concrete actions: a menu idea, music, an event and a partner cross-promotion. Each one cites the Qloo results it is built on, with a "watch out" note when a signal looks off | Agent over the evidence above + entity_tags |
| Ready-to-send kit | For every action: a WhatsApp broadcast in English and the city's local language (Kannada in Bengaluru, Spanish in New York, ...), an Instagram caption, specials-board text, a staff briefing, a partner outreach message and a checklist | — (written from the plan) |
| Poster | A ready-to-post 1080 × 1350 poster for the menu and event actions, with the headline in English and in the local language's own script (Kannada, Tamil, Hindi, ...), drawn in the browser; download or share straight to WhatsApp or Instagram | — (from the kit) |
| Crowd playlist | The music action comes with the artists this crowd loves, ranked by Qloo affinity, each one tap away on YouTube Music or Spotify | recommend (artists for the peer places' audience) |
| Your menu vs your crowd | Upload last month's item-wise sales (Petpooja or any POS CSV), or use the built-in sample month. Each dish is matched to a Qloo dish tag and scored for crowd fit, then sorted into Double down, Reposition, Rethink, Core staple or Keep, plus new dishes the crowd's favourite places are known for | find_tags, recommend (with the restaurant's peer places as taste signals), entity_tags |
| What-if | "Weekend biryani counter or a breakfast menu?" Ideas are scored for this crowd in one comparable run, with examples | find_tags, recommend |
| Competitor lens | "How is my crowd different from Vidyarthi Bhavan's?" Taste traits only places like yours have, only the competitor has, and both share | describe, compare_audiences |
| Generic chatbot vs TasteMenu | The same model answers the same question with no Qloo data, shown side by side ("0 real data points vs 13 Qloo signals"), so the difference Qloo makes is visible | — (baseline) |
Trust rules built into the code
- Every number on the board comes from Qloo. When the plan is published, each evidence item is matched by name against the Qloo responses of that session (
verifyPlan in src/agent.mjs). A matching item gets Qloo's own affinity value; a number Qloo did not return is removed; an unknown name is hidden.
- No gap-filling. If Qloo has no signal (no matching tag, competitor not found, empty comparison), the agent is instructed to say so instead of answering from general knowledge.
- Aggregate, not personal. Qloo results describe audience affinities, not individual customers, and the UI says so. Uploaded sales files stay in the session's memory and are never written to disk. Only the restaurant type and area are sent to Qloo.
How it works
flowchart LR
O[Owner, phone browser] -->|SSE stream| S[Node server<br/>src/server.mjs]
S --> A[Agent loop<br/>src/agent.mjs<br/>Gemini via OpenAI-compatible API]
A -->|tool calls| E[Evidence engine<br/>src/evidence.mjs]
A -->|tool calls| X[Sales x Taste, what-if,<br/>competitor lens<br/>src/sales.mjs]
E --> Q[Qloo bridge<br/>src/qloo.mjs]
X --> Q
Q -->|stdio MCP| H[qloo mcp<br/>official Qloo harness]
H --> API[(Qloo Insights API)]
Q --- C[(24 h disk cache)]
This is the beginning of the README. Read the rest at the original. Read the full README ↗
