All work

Multilingual document assistant

Ask in your language, get an answer with the page it came from.

Client
Swiss rail company
Type
Web app
When
2026
My role
Developer
Diagram: PDFs are chunked, embedded and stored in PostgreSQL; a question goes through hybrid search, reranking and a Gemini agent to an answer with page citations

Simplified system diagram. Client names, addresses and credentials are left out.

The problem

Staff and partners lose time searching long technical PDFs, often written in a language that is not their own.

What I built

Documents are split, indexed and searched with a mix of keyword and meaning-based search. An AI agent answers only from the passages it found, in German, French, Italian or English, and cites the document and page.

The result

Answers people can check in one click, instead of trusting a chatbot’s guess.

Built with

  • Python
  • PostgreSQL
  • Hybrid search
  • Gemini
  • Docker

Same work, four ways to see it.

Design
Typeface