Name in original language
CiudadanIA: Procesamiento con IA de Correspondencia Ciudadana para el Gabinete de Presidencia del Gobierno de España
Initiative overview
CiudadanIA is an AI initiative within the PresidencIA programme that applies generative AI to improve the processing of citizen correspondence addressed to the President of Spain through the official “Escribe al Presidente” (Write to the President) service. This channel receives thousands of letters from citizens raising concerns, making requests, and sharing opinions on policy matters. Each letter requires reading, summarising, categorising, and routing to the appropriate government department, a labour-intensive process that directly affects how quickly and effectively citizens receive responses.
CiudadanIA automates the summarisation step using small language models. Rather than relying exclusively on large cloud-based AI systems, the project combined cloud with fine-tuned compact models (specifically Gemma-3-4B, quantised for efficiency) to run on local government infrastructure. This design decision was driven by two requirements: citizen letters contain personal and sensitive information, and the government’s AI sovereignty objectives call for reduced dependency on third-party systems. The fine-tuning and evaluation process was rigorous. Government staff who process citizen correspondence daily provided expert feedback on model outputs, identifying systematic errors such as hallucinated information (inventing details not present in letters), tone distortion (characterising respectful requests as demands), and omission of relevant content. This feedback was used to iteratively refine prompts and model behaviour. To support this evaluation at scale, the programme developed an open-source LLM evaluation workbench that enables systematic testing of different prompts, models, and configurations.
The tool supports both cloud APIs and local models, includes batch processing for evaluating large sets of correspondence, and implements an LLM-as-a-judge methodology where one language model evaluates the quality of another’s outputs against configurable criteria (accuracy, conciseness, tone preservation). The tool also supports English and Spanish and is available for other government teams.
