AI Skincare Shopping Concierge — Full-Stack RAG Platform

What was built

Built a conversational AI shopping concierge for a D2C skincare brand, live in the UK and US, that gives personalized product recommendations backed by real-time retrieval instead of static content. Next.js 14 (App Router, TypeScript) frontend with Clerk auth, orchestrated by n8n and backed by Supabase Postgres with pgvector 0.8.0. Product/review/social data — 15,188 products (326 MB), 65,895 reviews (41 MB), and 11,654 social video entries (213 MB) — is indexed with OpenAI text-embedding-3-large (1536-dim) behind an HNSW index (m=16, ef_construction=64, cosine distance), queried via a hybrid vector+full-text-search RPC using reciprocal-rank-fusion. Three parallel agent tools power each query: catalog hybrid search (UK), live retailer scraping through Browserless for thin-coverage US retailers plus a catalog fallback, and a region-filtered discount-code lookup against Google Sheets. GPT-5.4-mini drives the conversation with prompt versions managed in Langfuse, returning product tables, social proof, reviews, and discount codes in one response. Deployed on Oracle Cloud with aaPanel, PM2, and Nginx.