-
-
TasteShift: know your crowd, choose your products, stock with confidence
-
Culture comparison: two 8,000-person events, live Qloo recommendations per audience
-
Live ops: mid-event restock, calibration and multi-day restock (simulated sales)
-
Three risk plans and what-if scenarios (rain, attendance, cost, rival stalls, MOQ)
-
Agent with vs without Qloo: every product decision cites Qloo films, audience age and taste tags
TL;DR
- Who: event vendors who must stock before the crowd arrives
- Problem: they only know the event name, so stock is a guess
- Solution: an agent asks Qloo who the crowd is and what else they like, then picks products. Every decision cites Qloo data
- Proof: vs the same agent without Qloo, it wins 17 of 20 runs, +NT$552 mean profit (95% CI +311 to +793), 92% of decisions cite verified Qloo data, 0 invented citations (simulated demand)
Try it in 1 minute
- Open https://demo.jxdtw.com/tasteshift/ (no login)
- Tab Agent (Qloo vs none) → Run both arms (about 1 minute)
- Compare: the left column cites Qloo for each product; the right column guesses
How the agent uses Qloo
| Step | Qloo API | What the agent learns |
|---|---|---|
| Resolve | find_tags |
event concepts → taste tags |
| Recommend | recommend, rank |
films this crowd loves, in Taipei |
| Profile | audience_demographics |
age and gender skew |
| Cross-domain | entity_tags |
food, brands, styles they also like |
Then the agent scores 12 products with cited reasons, and a deterministic optimizer sets quantities within budget. The LLM never sets numbers.
Example: what Qloo changed
- Qloo says: the Taipei anime-festival audience over-indexes at 35–44 (+0.60), under-indexes at 24 and younger (−0.19), and likes Cafe and Fashion
- Agent does: kids' sticker books ↓, band tees and tote bags ↑, each with its Qloo citation
Also in the app
Culture comparison of two events · 3 risk plans + what-if (rain, attendance, cost) · live restock · audit trail + purchase-order CSV
Built with
Python / FastAPI · Next.js · Qloo harness · function-calling LLM · 36 tests · nginx
Limits and next
- Demand is simulated → next: real point-of-sale data from vendors
- Code (MIT): https://github.com/seantw0301/tasteshift
Built With
- fastapi
- function-calling
- llm
- next.js
- nginx
- python
- qloo
- react
- tailwindcss
- typescript
