I build AI products people actually use, and prove they work.

QuedarBien

Most gifts end up in a drawer. Research on gift-giving says recipients value how much they use a gift, not the moment they open it. So QuedarBien asks what someone does on a normal day and what you always see them with, then finds the premium version of something they already use, at a price they'd never pay for themselves.

• The engine: LLM steps work out the objects a person uses every day and what a real upgrade looks like, hybrid search finds candidates in a 220,000-product affiliate catalogue, and one batched scoring call ranks them on use, upgrade and how known the brand is.
• Evaluated, not vibed: 20 realistic personas, answered the way real shoppers fill in forms (with plenty of "don't know"), scored by LLM judges each time the engine changes.
• Run as a product: per-visitor and daily cost caps so a search costs about four cents, cookieless ad attribution, and Meta ad tests to find the hook that works.

Under the hood: FastAPI with server-rendered pages and no client JavaScript; PostgreSQL with pgvector for hybrid (keyword plus embedding) search; product feeds from Spanish retailers via affiliate networks, refreshed by a script that swaps the production table; a rate limiter keyed on a salted visitor hash so no address is stored; deployed as a Docker container on my own Coolify server.

Version one (2025): Qubi, a chat-based gift genie on LangGraph, LangChain, FastAPI and AWS with managed PostgreSQL, built to learn agentic LLM workflows hands-on. It taught me to save LLM calls for what needs them: rule-based chat gathered preferences and code-based checks replaced model calls. The rebuild keeps that lesson and drops the chat for a short form people can actually answer.