Introduction
Ask for a product like you'd ask a friend.
Buyceps is an AI shopping assistant that accepts text, voice and image queries and returns ranked, personalized product recommendations — type a vibe, describe it out loud, or snap a photo of the thing you want.
It was built as a full-stack system: a Next.js storefront talking to a FastAPI backend that orchestrates the AI pipeline behind a single endpoint.
Architecture
Three agents behind one endpoint.
A CrewAI multi-agent pipeline splits the work: a vision-analysis agent interprets images, a product-search agent queries the catalog, and a response-generation agent writes the final recommendation — all coordinated behind one /agent/query API with parallel execution via thread pooling.
Search is hybrid: MiniLM sentence-transformer embeddings indexed in FAISS, fused with Llama-4 vision features, then ranked by multi-factor relevance scoring over an Amazon catalog.
3
cooperating CrewAI agents
3 modes
query by text, voice or image
FAISS
hybrid semantic index over an Amazon catalog
What's inside
Key features.
Multimodal input
Text, voice and image queries all resolve to the same recommendation pipeline.
Multi-agent orchestration
Vision, search and response agents cooperate through CrewAI with parallel execution.
Hybrid semantic search
MiniLM embeddings in FAISS fused with Llama-4 vision features.
Multi-factor ranking
Results scored on relevance signals, not just vector distance.
Single API surface
Everything behind one /agent/query endpoint the frontend can call simply.
Personalized output
Recommendations phrased and ranked for the individual query, not generic lists.
Built with