Featured project

Obrador Mel

A bakery with an assistant that prepares orders and a person who confirms them

Obrador Mel is the website of a fictional bakery in Santiago de Compostela, with a chat assistant that handles orders from start to finish: it reads the message in Spanish or Galician, works out the price, checks lead times and pickup slots and gets the order ready. The engine is written in JavaScript with no frameworks, and the model runs locally. The full case study is written in markdown in the repository itself and published on this portfolio.

Web

Source code Documentation Testing

Year: 2026
Role: Full-stack
Platform: Angular 22 · Node 24 · Ollama
Data: Postgres + Qdrant

The problem

An assistant that leaves everything to a language model can invent a price, or promise a cake for tomorrow that the bakery can't make in time. At a bakery, that mistake is an order nobody can fill. Even when it gets things right, nobody at the shop has seen what was promised.

The solution

Most turns are solved without the model, in under 30 ms: the engine understands the message and works out prices, lead times and slots from the business's own data. The local model only writes answers from the sources the engine gives it; if a reply has a price, a lead time or an hour that isn't in them, it's thrown out and the source text is used instead. The order reaches the staff panel, where a person confirms it, rejects it or proposes another time, and the customer sees that in the chat. If the model or the vector database go down, the demo keeps working. 175 tests and 32 end-to-end conversations check all of this.