CiudadanIA summarises citizen letters to the Spanish President with small language models on local government infrastructure

Overview

The Spanish Prime Minister's Office uses fine-tuned compact models (Gemma-3-4B, quantised) running on local government infrastructure to summarise citizen letters sent through the 'Write to the President' service, supporting routing and response. Government staff reviewed outputs for hallucination and tone distortion, and the project produced an open-source LLM-as-a-judge evaluation workbench. Humans stay in the loop; the source lists impacts qualitatively, without figures.

The challenge

The Spanish President's 'Write to the President' service receives thousands of citizen letters, each of which needs reading, summarising, categorising and routing to the appropriate government department, a labour-intensive process that affects how quickly citizens receive responses.

The solution

CiudadanIA automates the summarisation step using small language models. The project combined cloud with fine-tuned compact models (Gemma-3-4B, quantised for efficiency) that run on local government infrastructure, because the letters contain personal and sensitive information and the government's AI sovereignty objectives call for reduced dependency on third-party systems. Government staff who process the correspondence gave feedback on outputs to refine prompts and model behaviour, and the project keeps a human in the loop.

Reported business value

The source lists impacts qualitatively, without figures: faster processing of citizen correspondence, improved service quality, reduced staff workload on routine summarisation, and a reusable open-source evaluation methodology for other government AI deployments.