About

How I got here

I didn’t start out planning to become a data scientist. I studied agronomy engineering at Institut Agro - Agrocampus Ouest (Rennes) — which is fundamentally about applying science to complex living systems — and during my statistics courses I realized that the analytical side was what excited me most. Not just running models, but understanding why a method works, and what it can and can’t tell you about the real world.

After finishing both my engineering degree and a master’s in applied mathematics and statistics in 2021, I joined Astek / IT&M Stats, a consultancy that places statisticians and data scientists inside large R&D organizations. That structure has shaped how I work: every 12–18 months, a new domain, a new team, a new set of constraints. It forces you to get good at context-switching, at earning trust quickly, and at building things that work without you being there to explain them.

My first mission was at L’Oréal R&D — two years of clinical biostatistics on ingredient efficacy and molecular safety studies. I covered hundreds of analyses, learned what a 2-week turnaround really means under pressure, and built internal tooling that helped the team move faster. It was a good introduction to regulated science: the kind of work where you can’t hand-wave uncertainty.

Then Sanofi R&D, where I shifted from pure analysis to building. I designed and shipped a full R Shiny platform for managing pilot manufacturing chains across 100+ clinical projects — took on the Scrum Master role alongside the development work. That was the project that taught me that useful software is harder to build than correct software.

After Sanofi, a stint at Abolis, a microbiome biotech startup. Six months, full autonomy, multi-omics data pipelines, no hand-holding. Very different energy from a pharma company, and I learned a lot from it.

Now I’m at Chanel Parfums Beauté R&D, working on machine learning and computer vision for fragrance and cosmetics development. Still on mission via Astek/IT&M Stats, but this is the most technically ambitious role so far — models that go into production, tools that scientists depend on daily.

Outside of work: I’ve been following One Piece for years — not casually, properly following it. It’s one of the few long-form narratives that actually earns its length, and I find the way it builds on itself across decades genuinely interesting. I play video games when life allows, which is less often than I’d like. And I have a lot of time for people who can say “I was wrong about that” without making a production of it — the ones who just update and move on. It’s rarer than it should be.

What I care about at work

Working at the interface, not the edges. The most useful position is between the domain experts (chemists, clinicians, biologists) and the data/software systems. People who sit purely on one side often build the wrong things. My agronomy background genuinely helps here — I can read a protocol, understand an experimental design, and push back when the question being asked can’t be answered with the data available. I find these cross-functional, partner roles genuinely more interesting than working in isolation.

Tooling that disappears. The best internal tools are the ones that become invisible — people just use them. A dashboard that requires a walkthrough every time is a failed tool. I spend a lot of time thinking about user experience (UX) in contexts (scientific apps, internal platforms) where it’s usually ignored.

Rigor over speed, mostly. Working in pharma and cosmetics R&D means results matter. Not in an abstract “good science” way — in a “this goes into a regulatory submission” or “this product formula ships to production” way. I’ve developed a healthy caution about overfitting, about p-values, about claims that outrun the data. The constraints of regulated industries aren’t obstacles; they’re clarifying.

Understanding models, not just using them. I’m drawn to the deeper mathematical and statistical machinery underneath the tools I reach for daily. Not for its own sake, but because I’ve found that understanding why a method works — its assumptions, its failure modes, its relationship to other approaches — makes you better at the judgment calls that actually matter.

Explainability and the limits of black boxes. In regulated R&D, a model you can’t explain is often a model you can’t use. But beyond compliance, I’m genuinely interested in interpretable machine learning (ML) and AI as a research area — not as a constraint to work around, but as a more honest representation of what we actually know.

Staying current without being credulous. I keep a fairly active scientific watch (veille scientifique) — following the methods literature in biostatistics, computational biology, and ML engineering. I try to distinguish between tools that have earned their place and things that are interesting but not yet ready. The field moves fast enough that this requires active effort.

Open source tools and reproducible workflows. R, Python, Quarto, Git, Docker. Reproducibility is not just a technical preference — it’s a practical requirement in regulated science, and it’s how you avoid the slow-motion disasters that come from analyses nobody can reconstruct. I’ve seen what happens to work that wasn’t built this way.

Get in touch

Content on this site is licensed under CC BY NC SA 4.0 unless stated otherwise.

Back to top