In this interview, Hubadviser had the opportunity to speak with Thomas Ménard, Deputy Head of IT Operations at the Institut Pasteur, who explained how the IT department has worked to enable business teams to harness the power of AI securely, without generating frustration or encouraging shadow AI practices.
We discussed key concepts such as self-service, LLMs, and RAG, as well as the infrastructure challenges involved.
Generative AI in the service of eesearch
A demanding scientific environment
Ismail Charkaoui : The Institut Pasteur is globally recognized for its advances in biology and medicine. But managing scientific knowledge is a real challenge. How does generative AI fit into this context?
Thomas Ménard : Exactly ! Scientific research relies on an enormous amount of data, and one of the main challenges is how to structure and leverage it effectively. The arrival of generative AI has opened up new perspectives, but also raised governance issues. Very early on, we asked ourselves how we could provide our employees with a secure tool, to prevent them from using ChatGPT and potentially sharing sensitive data.
Ismail Charkaoui : So the idea was to offer an internal, controlled alternative ?
Thomas Ménard : Absolutely. Our goal was threefold:
- To protect the confidentiality of scientific and medical data
- To centralize and regulate the use of AI in order to get the most out of it
- And to encourage the adoption of generative AI across teams so we wouldn’t fall behind competitors
Ismail Charkaoui : I see. What were your options for allowing the use of generative AI internally ?
Thomas Ménard : The easiest even obvious option was to turn to Microsoft Copilot licenses. We tested the solution but weren’t convinced by its user experience. Several teams told us directly they preferred using their personal ChatGPT accounts over switching to Copilot, which is a major security risk.
On top of that, the pricing for the Copilot versions we found interesting was too high. It would have forced us to juggle between a limited number of licenses and growing demand for generative AI. Our OPEX could have exploded. So after some consideration, we ruled that option out.
Ismail : And the solution you chose was LibreChat. Can you describe it and explain why?
Thomas : LibreChat is an open-source chatbot platform, with a user interface very similar to ChatGPT. Internally, we can:
- Customize the UI (logo, colors, usage rules)
- Freely choose the AI model(s) behind LibreChat (LLaMA, Mistral, etc.)
- Secure and control access via centralized authentication (LDAP, SSO) and encrypted communication
- Connect LibreChat to various internal data sources (APIs, document repositories, internal databases)
Ismail : Why did you choose this solution?
Thomas : LibreChat stood out for several reasons :
- Its intuitive UI very close to ChatGPT’s, which helps boost adoption
- Its modularity, which allows us to integrate multiple LLMs
- Its self-hosting capabilities, which ensure data sovereignty
- Its active open-source community, constantly improving the tool
Ismail : You mentioned using several LLMs can you explain which ones and why?
Thomas :
We currently use about ten models, ranging from small, fast models to larger, GPU-intensive ones. Among them: OpenAI, Mistral AI, LLaMA, and even DeepSeek, which we recently tested.
We adopted a bimodal approach:
- Cloud-based models (e.g., OpenAI, Anthropic)
✅ High performance
❌ Risk of data leakage - On-premise models
✅ Maximum security
❌ Expensive infrastructure and more complex maintenance
The goal is to let users choose the most appropriate LLM based on the use case. But this requires education and guidance, because it’s not always easy for users to understand which model to choose.
Eventually, we want to create a centralized hub with:
- Simplified access to both on-prem and cloud models
- A unified API to integrate AI into business workflows
- Monitoring and traceability tools (usage logs, dashboards)
Ismail : So the idea is that every employee can use AI without dealing with the technical complexity?
Thomas : Exactly. We want LibreChat to become a single point of access for generative AI secure and optimized.
Ismail : When I visited your offices, I saw a tool that could answer specific questions using Pasteur’s confidential data a powerful management tool, capable of mapping out all ongoing vaccination projects across Pasteur. That was previously impossible. Can you explain how that works?
Thomas : What you saw is a SQL agent. It converts a natural language question into an SQL query and then interrogates an internal database. It’s perfect for real-time stats and decision-making.
Ismail : So this is part of your self-service approach?
Thomas : Exactly. Everyone talks about AI, but less about data. Yet AI can help us become more data-driven. Thanks to LLMs, querying becomes much easier just natural language. This makes data-driven decision-making more democratic. It removes the middle layers and allows teams to access the data they really need.
Ismail Charkaoui : Is this SQL agent based on RAG ?
Thomas : Not exactly. Unlike the SQL agent, RAG (Retrieval-Augmented Generation) pulls information from unstructured documents (reports, memos, publications). But it requires rigorous document governance, and we’re not fully there yet.
Ismail Charkaoui : What are the main challenges with RAG ?
Thomas Ménard : Data quality. Many companies think a RAG system can “magically” structure their knowledge, but if documents are poorly organized or outdated, the results will be disappointing.
Ismail Charkaoui : Did you bring in an external partner for this ?
Thomas Ménard : Yes structuring a document corpus is a real profession. They helped us :
- Map existing data
- Define structuring and update rules
- Optimize document ingestion pipelines
- Today, our internal RAG system is progressing well.
Ismail : To advance this far with AI, you need strong infrastructure. How did you manage it?
Thomas : That’s right. We already had a GPU cluster and a containerized architecture (Docker, Kubernetes).
This allowed us to deploy LibreChat quickly and add new models over time. That’s because Pasteur has always invested in its IT infrastructure. We have the computing power and a clear infrastructure strategy, which gives us an edge in launching these kinds of projects.
Conclusion
This case study from the Institut Pasteur highlights both the opportunities and challenges of adopting generative AI in a demanding professional environment. By choosing LibreChat, Pasteur was able to ensure data sovereignty while providing employees with simplified access to a variety of LLMs. Beyond the internal chatbot, the implementation of a SQL agent also demonstrates how AI can boost decision-making and performance management by democratizing access to data.
However, this enthusiasm for AI should not obscure a key reality: the success of any AI project depends as much on the quality of the infrastructure and robustness of IT processes as on documentation governance. The example of RAG (Retrieval-Augmented Generation) illustrates this clearly it requires true maturity in information management and structuring, justifying the support of external experts.
In the coming months and years, the Institut Pasteur will continue to expand its AI hub, combining on-premise models with cloud-based solutions. This approach reflects the need for balance: leveraging the power of generative AI without compromising security or data sovereignty.
For organizations like Pasteur, aiming to unlock the value of complex document assets, the lessons learned will serve as a guide and pave the way for further innovation. These are the kinds of innovations Hubadviser is committed to tracking and sharing.
About Thomas MÉNARD
Thomas Ménard is Deputy Head of IT Operations at the Institut Pasteur. With over 25 years of experience in IT, he has held roles in major organizations such as Amadeus, Bouygues Télécom, and Institut Pasteur. Now Deputy Head of Production, he leads a team and works in highly technical environments. He has contributed to the adoption of DevOps, Kubernetes, and Pasteur’s microservices strategy. His current focus is on integrating AI solutions to improve operational efficiency.
