Enhancing Public-Sector Decisions Through Specialised AI

Organisations involved

Main Participant: Delibia is a French SME developing generative AI tools that help local authorities analyse and compare their decision-making processes.

ISV: The Occitanie Region supplies public data, pilots real-world testing of the platform, and trains civil servants to use the solution.

HPC Provider: GENCI provides High-Performance Computing resources via the Jean Zay supercomputer, enabling large-scale model training and optimisation.

The challenge

Local authorities in France publish millions of governance documents every year, but these are made available across thousands of websites for public reference and scrutiny. Such extensive fragmentation makes it difficult for officers and elected representatives to compare how similar issues have been managed in other regions. Before approving a project or implementing policy, they must understand the options chosen by comparable regions, the constraints and reasoning behind each decision, and the outcomes achieved. Without structured access to this information, decision-making is slower, less informed and inconsistent across the country.

Delibia aims to convert this content into actionable knowledge via a search platform built for the public sector. The challenge is that existing generative AI models are designed for broad use cases. They struggle with long administrative text, lack the legal and governance precision required, and do not provide reliable traceability—key requirements in public sector administration. Creating a specialised model demands the processing of millions of documents and successful optimisation of hundreds of millions of parameters. These workflows require tens of thousands of hours of GPU compute, well beyond the reach of an SME limited computational resources. Consequently, access to large-scale HPC resources was essential to develop an accurate, long-context model that meets the needs of regional and municipal Government.

The Solution

Delibia developed a long-context generative AI model tailored to public-sector documents. Using GENCI’s High-Performance Computing resources, the team fine-tuned language model representations (known as embeddings) on more than three million government administrative documents (referred to as administrative acts) requiring over 50,000 GPU hours.

The resulting model is four times smaller and delivers 2.3-times faster responses, reducing memory and compute requirements for local authorities. It offers more precise retrieval, better understanding of administrative language, and full on-premise deployment without external services, ensuring data sovereignty and compliance with strict public-sector security standards.

Impact 

The new embedding model improves the accuracy and speed of retrieval, legal references and regulatory clauses. Tasks that previously took minutes can now be completed in seconds, reducing the workload for officers and support teams. For Delibia, this strengthens its SaaS platform and enables premium features focused on public-sector workflows. Based on current interest from regions and major cities, these improvements could generate a 15–20% rise in recurring revenue over two years. The ability to train sovereign, domain-specific models also reinforces Delibia’s competitive position.

Socially, the platform improves transparency and empowers local government officers by making millions of public decisions easier to compare and understand. It supports faster, more informed decision-making and governance across territories.

Environmentally, the compact model requires far fewer computing resources, reducing energy consumption and contributing to more responsible digital operations.

Benefits

• 2.3× faster inference and 4× smaller footprint, reducing response times and energy consumption.
• High-accuracy retrieval across millions of administrative acts, improving daily efficiency for officers and legal teams.
• Sovereign model enabling GDPR & ISO 27001 standard compliance.
• Lower infrastructure costs thanks to a compact model that runs on a single GPU server.
• Reusable training methodology for sector-specific embedding models, which can be extended to other public-sector domains.