ThinkDeep's DeepBrain agentic platform cut French government document processing from two days to two minutes and saved the Ministry of Economy and Finance €2 million at scale across 10,000 employees, with each 100-user deployment recovering an additional 1,000 work hours per month.
"ThinkDeep AI knowledge and expertise have been instrumental in deploying and scaling our generative AI platform," said Thomas Binder, head of AI at the French General Directorate of Public Finances, in NVIDIA's published case study.
The French administration annually handles more than 100 million documents to provide tailored responses to citizen inquiries, with public servants required to cite specific legal references and stay current with changing rules while keeping every byte of citizen data inside government infrastructure to comply with the European AI Act. ThinkDeep built DeepBrain specifically to meet that on-premises sovereignty requirement.
The rollout moved from a one-week DeepBrain instance setup to a 10,000-employee production deployment inside the French Ministry of Economy and Finance, with a parallel deployment covering the Ministry of Defense. Over that span, document-retrieval time compressed from two days to two minutes and operational cost savings compounded as seats scaled.
Under the hood DeepBrain runs on NVIDIA DGX H100 and DGX H200 GPUs deployed on-premises inside the ministry's data centres, with NVIDIA AI Enterprise software orchestrating the model layer. NVIDIA NIM packages and deploys the latest LLMs on-prem, NVIDIA Dynamo optimises inference speed for reasoning models, and the NeMo framework integrates LLMs and vision-language models. NeMo Retriever grounds responses across millions of files while NeMo Guardrails enforces safety against biased or harmful outputs. The DeepBrain multi-agent layer routes tasks to specialised agents for fraud detection, legal document processing, security analysis and code review, and the DeepBrain Graph methodology enables real-time, sourced search across millions of files. NVIDIA Jetson Orin handles edge installations and NVIDIA Llama Nemotron models power the French-language reasoning layer.
In production across both the Ministry of Economy and Finance and the Ministry of Defense, DeepBrain serves thousands of public servants per ministry, processing millions of PDFs, scanned images, schemas and videos per day with response times in the two-minute range and energy-management savings from controlling the model layer in-house. ThinkDeep cites traction beyond the French public sector in finance, aerospace and energy — all regulated industries facing similar sovereignty constraints.
The ThinkDeep–NVIDIA partnership is now evolving toward deeper multi-agent AI reasoning systems, with ThinkDeep exploring applications in finance, healthcare and logistics where the same sovereign-agent pattern applies.