Lyon holds one of Europe's most concentrated life-sciences clusters, a chemical industry running down the Rhône, a broad manufacturing base, and the logistics that follows from sitting on the crossroads between northern Europe, Italy and the Mediterranean. Four industries, and in practice two kinds of project.
The first is document extraction where being wrong is expensive. Study reports, regulatory submissions, batch records, deviations, safety data sheets, supplier declarations. The naive version of this project — put the documents in a vector database and ask questions — produces a demonstration that impresses a steering committee and fails the first audit, because it cannot show where an answer came from and cannot say when it does not know. The version we build cites the page, refuses more often, and routes the uncertain cases to a named reviewer. That pattern is set out in clinical document intelligence, and the research-facing variant under life sciences research.
The second is operational: sensor histories, maintenance logs and quality records from plants that have been instrumented for longer than most software companies have existed. The data is there, the labelling usually is not, and the honest first step is often a month of establishing whether a signal exists at all before anyone commits to a model. Predictive maintenance and supply chain planning describe how those run.
Both kinds share a trait worth naming: the person who understands the data is a chemist, a QA lead or a scheduler rather than a data engineer, and the project succeeds or fails on how much of their time we get. We would rather have six hours with them than six weeks of documentation.
