“In God we trust. All others must bring data.”
— W. Edwards Deming
I am interested in understanding biological mechanisms, and in the statistical work that makes those questions answerable: how diseases progress, how treatments act, why patients respond differently, and which models can reflect the underlying biology closely enough to separate signal from noise.
I first joined Roche (Basel Headquarters) as a research intern, applying latent variable models to disability endpoints in neuroinflammatory diseases, and returned full-time to a translational and exploratory biomarker team, where my current focus is haemophilia and gene therapy. I use molecular and clinical data to characterise mechanisms of action, understand response heterogeneity, and identify biomarkers that help guide the development of new therapies. I also support clinical studies as a trial statistician, contributing to study design, interim analyses, and decision rules.
Across both, the aim is evidence that holds up where it matters: in decisions about which patients, which endpoints, and whether to continue.
Before that, I was a postdoctoral researcher at the Clinical Research Center of the Geneva University Hospitals (HUG), where I led multi-study meta-analyses of survival data across European clinical registries, developing inference strategies for heterogeneous data. In parallel, I was the main statistician on longitudinal cohort studies of early predictors of Alzheimer’s disease and protective factors against amyloid accumulation. I hold a PhD in Statistics from the University of Geneva (co-advisors: Prof. Maria-Pia Victoria-Feser and Prof. Stéphane Guerrier), where I developed new methods for bioequivalence testing, designed a generative AI model for spatial extremes, and contributed to simulation-based privacy-preserving inference techniques.
Overall, my work bridges statistical innovation and practical impact, with over seven publications, ten talks, and open-source tools that foster reproducible and collaborative research in statistics and machine learning.
Research included:
Thesis: Learning Max-stable Distributions with Generative Adversarial Networks. Advisor: Prof. Sebastian Engelke.
Grade: 5.5/6
A unified framework for multivariate bioequivalence assessment of locally acting drugs, providing size control and improved power under realistic conditions.
We propose the multivariate $\alpha$-TOST, a finite-sample correction that preserves size and improves power in multivariate (bio)equivalence studies.
We introduce the $\alpha$-TOST, a finite-sample correction to the TOST procedure that restores nominal error rates and improves power in (bio)equivalence studies.
We introduce evtGAN, a hybrid EVT–GAN model that better captures spatial climate-extreme dependencies than standard GANs or classical EVT methods.
An 8min overview of QRPs, why they matter, and what they look like in a real clinical trial protocol.
This talk presents improved finite-sample corrections for equivalence testing, offering better calibration and power than standard methods.
A short overview of why the EDSS is limited and how Item Response Theory might provide a more sensitive approach to modeling disability progression in MS.
An introduction to DP-JIMI, a simulation-based approach to inference under differential privacy.
A quick, example-driven overview of differential privacy and how mechanisms like Laplace noise provide strong privacy guarantees with useful accuracy.
A concise introduction to Edgeworth expansions and how they refine asymptotic approximations in statistics.
A hybrid framework combining extreme value theory with GANs to learn and simulate high-dimensional spatial extremes.
A concise introduction to how GANs learn data distributions through adversarial training and why they’re powerful yet hard to train.
evtGAN is a lightweight emulator that combines extreme value theory (EVT) with generative adversarial networks (GANs) to model rare compound events with far greater accuracy than traditional tools. Climate models are computationally costly, EVT alone struggles with complex spatial dependence, and standard machine-learning models typically fail in the extreme tail. evtGAN overcomes these limitations by separating marginal behavior from dependence using a copula-based framework: EVT provides theoretically sound marginal modeling and tail extrapolation, while GANs flexibly learn spatial dependence patterns. With strong performance even from as few as 30 annual maxima, evtGAN offers scientists an efficient, ready-to-use solution for simulating extremes. The Zenodo repository provides 2'000 years of simulated annual temperature and precipitation maxima over Western Europe, along with R and Python (TensorFlow) code to reproduce the method from our Environmental Data Science (2022) publication.
Tell me about your project, collaboration idea, or question.