Useful assistance starts with the service task
Health-insurance service teams answer questions that depend on detailed organisational knowledge. An email may require finding a relevant procedure, understanding the request, and preparing an appropriate response. Automating part of that work means fitting AI into the service process, with a clear view of where staff expertise remains necessary.
During his 2023–2024 role at CM, Gilles developed an AI-driven system for client-facing and internal email correspondence. He also worked on an LLM-based virtual assistant and engineered its retrieval-augmented generation backend, alongside structured-data science and ML projects. This page describes those contributions within the CM team.
Bring organisational knowledge into the answer
A retrieval-augmented assistant connects a language model to an information layer relevant to the user's question. The backend is where that connection becomes an application capability: receive a request, retrieve useful material, and provide context for the conversational response. Its quality depends on both information access and the behaviour of the model using that information.
That distinction is particularly useful in service environments. A missing answer may reflect missing source material, weak retrieval, an ambiguous request, or a generation error. Separating those possibilities gives a delivery team concrete questions to investigate rather than repeatedly rewriting a prompt without understanding the cause.
Correspondence and conversations share a foundation
Email assistance and a conversational interface present information differently, but both need a sound connection between the request and the knowledge used to respond. The documented work spans those two forms of assistance. It also draws on a broader data-science remit, where structured information and language tasks can contribute to the same operational problem.
For a system of this kind, implementation choices should follow the workflow. Which requests are suitable for assistance? What information is needed to answer them? Where should a staff member review or take over? These questions provide a practical frame for connecting the AI component with existing service responsibilities.
Build capability around a real operating need
The contribution includes development and implementation of correspondence automation and the engineering of an LLM assistant's RAG backend. Both connect language technology with a concrete service need. The broader data-science work adds another perspective: an organisation's AI opportunities span the information in messages as well as the information in structured records.
The case shows why enterprise AI benefits from a concrete starting point. Focus on a recurring task, connect the model to the relevant information, and make the application useful to the people responsible for the service. The result is a more coherent engineering objective than a general request to add AI across an organisation.
