Talks
My conference presentations and poster sessions at academic venues.
Conference Presentations
Just-in-Time soft sensors fit a local model for every new batch, but they struggle with hundreds of drifting process variables. I presented our accepted paper, which retrieves similar batches in the latent space of a Self-Organizing Map (SOM) rather than in the raw sensor space, and keeps the model interpretable with an online SHAP-based feature selection that tracks the most influential variables as the process evolves. On a Michelin rubber production line, it cuts the mean squared error by up to 61% and lifts R² from -0.14 to 0.56 over global baselines. Presented at the 31st IEEE International Conference on Emerging Technologies and Factory Automation, a premier conference on industrial automation sponsored by the IEEE Industrial Electronics Society.
A week-long gathering of France's young statistics community, organized by the French Statistical Society (SFdS), bringing together PhD students and early-career researchers for talks, tutorials, and exchange. I presented Self-Organized Conformal Prediction, my work on closing regional coverage gaps in uncertainty quantification through unsupervised group discovery.
Predicting product quality in real time is hard when the manufacturing process itself keeps drifting. This work benchmarks five adaptive soft sensing strategies, from temporal weighting to neighborhood-based retrieval, leveraging Self-Organizing Maps (SOMs) with online SHAP-driven feature selection (SHAP scores how much each variable drives a prediction). Evaluated on 35,125 production batches and 167 process variables at Michelin, it gives practical guidance for choosing the right online adaptation strategy in industrial settings.
Rubber compounding involves hundreds of interacting process variables, and only a few actually drive product quality. This work takes an explainability-first approach, combining gradient-boosted trees (a strong tabular model built from many small decision trees) with SHAP attribution and conformal prediction, giving engineers both interpretable insights and statistically rigorous uncertainty estimates. The result: process experts can pinpoint critical quality drivers with quantified confidence. Presented at ESCAPE 35, one of Europe's premier conferences in computer-aided chemical engineering.
Posters
Standard conformal prediction gives prediction intervals that are correct on average, which can hide undercoverage in specific, safety-critical process regimes. Self-Organized CP closes that gap by discovering groups directly from the input geometry with a Self-Organizing Map (SOM), then calibrating locally inside each one, with no supervised labels and no model retraining. I presented this poster at two venues: SIAM UQ26, a leading conference on uncertainty quantification for industrial applications, and Michelin's annual doctoral day.
A black-box quality prediction is only useful once a process engineer can act on it. This poster presents a unified framework pairing gradient boosted models with SHAP attribution and conformal coverage guarantees, translating raw predictions into interpretable process insights at production scale.
Awarded Best Poster in the Mathematics category.
The first iteration of my knowledge discovery framework, focused on offline analysis. Gradient boosted regression combined with SHAP attribution and conformal prediction intervals surfaces the most influential process variables and their interactions, giving rubber compounding engineers a clear, quantified view of what drives product quality.
Thin line features in noisy images are notoriously hard to detect with sub-pixel accuracy. The 2DSBG filter uses an asymmetric Gaussian kernel that selectively enhances elongated structures while suppressing background noise, outperforming classical symmetric approaches like the Laplacian-of-Gaussian on both synthetic and real-world benchmarks. Presented at ICASSP, the flagship IEEE conference on signal processing.