ThinkDeep's DeepBrain AI Agents Save French Government €2M, Cut Document Retrieval from 2 Days to 2 Minutes

The French Ministry of Economy and Finance saved €2 million by rolling out ThinkDeep AI's DeepBrain multi-agent platform to 10,000 employees, according to an NVIDIA customer story. Retrieving relevant legal information, which used to take up to two days, now takes two minutes, and a typical 100-user deployment saves about 1,000 work hours a month. The Ministry of Defense is also a customer. "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. The French administration handles more than 100 million documents a year to answer citizen inquiries, and public servants must track changing regulations and cite the exact legal references for each case. The work involves sensitive data and must comply with the European AI Act, so the system had to run inside the government's own infrastructure. DeepBrain lets public servants build their own domain-specific agents connected to many data sources, including PDFs, scanned images, schemas and videos. The agents consolidate that material into knowledge bases and cross-reference it, and a "DeepBrain Graph" method supports searching across millions of files and their relationships while returning sourced answers. Users can run, for example, a fraud-detection agent or a legal-document agent built on aggregated French legal documentation, and agents can call one another for cross-domain workflows. Data stays on the customer's infrastructure at all times. ThinkDeep built the platform with NVIDIA H100 and H200 GPUs and NVIDIA AI Enterprise, using NIM microservices to package and run LLMs on premises, NeMo Retriever, Llama Nemotron models, NVIDIA Dynamo for inference speed, the NeMo framework to bring LLMs and vision-language models into existing infrastructure, and NeMo Guardrails against biased or harmful outputs. The finance ministry chose NVIDIA DGX for its on-premises hardware. Setting up a DeepBrain instance and its connectors took about one week, and running its own models and hardware also let the ministry manage energy costs.