Featured work

Elections are the most sensitive domain there is for AI. Voters base their vote on the information they find, so a single misrepresented party position could damage trust in the voting aid and in the broadcasters behind it. The bar was therefore uncompromising: neutral, verifiable and factually accurate, at the scale of hundreds of municipalities and more than 1,300 local parties.

An advanced RAG pipeline that answers exclusively from official party programmes, with a clickable source citation on every answer. Smart retrieval technology (query expansion and semantic reranking) understood questions in everyday language, and per-party retrieval gave the smallest local parties an equal place next to the big ones. On top of that we built a multi-tenant admin panel through which every broadcaster and political party managed its own documents, over 1,300 parties across nearly 240 municipalities.

Build the Netherlands’ first AI voting assistant. RTV Rijnmond was among the first newsrooms in Europe to embrace AI in election journalism, and that pioneering role demanded a solution in which the newsroom kept full control over every answer. There was no playbook: we had to prove it could be done.

A hybrid architecture: an NLP engine (Rasa) recognised the topic and entities in every question and matched them to editorially approved answers. Only when no editorial answer existed did an LLM fallback take over, clearly marked as such. Maximum control for the newsroom, maximum coverage for the user.