Seminars
My seminar presentations and paper discussions.
Reading and debating research is central to how I work. This page gathers the seminar I founded at Michelin, a selection of papers I have presented there, and talks I have been invited to give.
A Research Seminar I Founded
A Paper A Week (APAW)
I founded and led a bi-monthly research seminar that brought together 20+ researchers and data scientists across Michelin. Each session dove into a single paper or topic, from conformal prediction and Self-Organizing Maps to long-range sequence models and the transformer architecture, always bridging the latest academic ideas and the industrial problems we faced daily. The goal was simple: keep learning together and spot opportunities to bring cutting-edge research into production.
April 2024 to January 2026 · Clermont-Ferrand · Bi-monthly · 20+ Researchers and Data Scientists
Selected Paper Presentations
A sample of the papers I have presented at A Paper A Week, chosen for how they connect frontier research to the problems we face in industry.
Presented my own open-source library for Self-Organizing Maps in PyTorch, showing how SOMs can power industrial data visualization, anomaly detection, and similarity-based retrieval at scale.
A sub-quadratic alternative to transformers for extremely long sequences (up to 1M tokens). I presented this for its potential connections to industrial time-series modeling where sequences can span thousands of process steps.
An accessible introduction to conformal prediction, a framework for producing prediction intervals with finite-sample coverage guarantees. Directly relevant to my PhD work on uncertainty quantification for industrial quality monitoring.
An overview of QD algorithms, which find not one but an entire repertoire of diverse, high-performing solutions. I presented this from my MSc experience at Imperial College, discussing potential applications to process optimization.
The foundational paper behind modern language models, vision transformers, and much of today's AI. I discussed the self-attention mechanism and its implications for sequential industrial data, foundational background for most of modern deep learning.
Invited Seminars
Talks I was invited to give to research and data-science audiences.
After our 2nd place at Hack the World(s), the ORAILIX team of LIX, where my teammate Lucas Thil is a PhD student, invited team Piaget to present the project. We shared the motivation behind a JEPA world model built on a complete fly-brain connectome, the technical challenges we tackled, such as the 96% neural silence that pushes JEPA toward collapse, and the results we obtained.
Presented TorchSOM to the mathematical PhD community at Polytechnique, covering the library's design, its role as a backbone for Just-in-Time Learning, and applications to online quality sensing and conformal prediction in industrial settings.
Presented my work on adaptive soft sensing to Michelin's Data Scientists Network, comparing temporal, similarity-based, and hybrid strategies for real-time quality monitoring on the production line.