Case study
Radio Perico
A translator with a very specific voice.
Paste ordinary Spanish and it comes back in the voice of Perico Delgado, the Tour winner turned cycling commentator. Live at radioperico.com, with a public course on how it was built.
RAG over a corpus of his own words
A corpus built from transcribed broadcasts and classified with an LLM, embedded with a multilingual sentence-transformers model and indexed in pgvector on Supabase. Each request retrieves the closest phrases and hands them, as examples to follow, to the Claude API — with the right model for each use case.
A FastAPI backend in Python
Python and FastAPI serve the translation as a stream, with embeddings from Hugging Face inference and pooled connections to Postgres; React and Vite on the front, with dictation and speech synthesis. An evaluation suite and spending caps keep it honest.