What it does
PayFox is an AI commerce agent that grounds product recommendations in real cultural taste data. Tell it what you're into — music, films, hobbies, aesthetics — and it maps you through Qloo's Taste Graph (250M+ entities) to brands you'd actually love, then closes the loop with PayPal checkout in the same conversation.
The problem
LLMs can recommend anything. What they can't do is know what people actually like. Ask a generic model for brands that fit "Wes Anderson + vinyl records" and it guesses. PayFox doesn't guess — every recommendation is anchored in Qloo's cultural affinity data, the same graph that powers recommendations for major global brands.
The demo
Four one-tap personas (Jazz + Noir, EDM + Gaming, Indie + Nature, Wes Anderson) each run a live three-step Qloo pipeline:
qloo_search → qloo_insights → taste_recommend
Every result shows the honest side-by-side: what a generic LLM would say vs. what Qloo's graph knows — then taste-matched brands, one tap to buy via PayPal.
How we built it
- Qloo API: Search finds cultural entities from natural language; Insights returns brand affinity based on entity IDs as signals. The agent calls these as tools.
- Muse Spark 1.3 (Meta Model API): the agent brain — routes intents, calls tools, writes every reply (zero hardcoded responses).
- PayPal Orders API v2: checkout from the recommendation cards. Same agent, dual hackathon entry.
- Go backend + web chat UI, hosted and live.
The insight
Recommendation agents fail when they treat taste as trivia. PayFox treats taste as data — 250M entities of it — and the "without Qloo vs with Qloo" comparison is the product, not the pitch. Remove Qloo and the recommendations collapse into LLM hallucination.
Challenges we faced
- Teaching the agent to route personas through the right Qloo endpoints (Search → Insights) instead of answering from the LLM alone — solved with explicit tool descriptions and intent routing in the system prompt.
- Extracting individual taste entities from freeform text ("I love Miles Davis, jazz, and French noir films" → three separate Qloo searches) so the taste signals are clean, not one messy blob.
- Keeping checkout inside the chat without breaking the recommendation context — PayPal is wired as another tool in the same agent loop.
Built With
- agentic-commerce
- facebook-messenger
- go
- meta-model-api
- muse-spark
- paypal-orders-api
- qloo
- taste-ai
