CV

Real-time Spatial AI for retail: open-world computer vision, with Sacha Hu, Ashley Napier, and Galen Han.

• Developed open-world VLMs that recognize products, aisles, and price tags from images alone, trained on paired images and product descriptions for identification and planogram matching.
• Used test-time augmentation to estimate the confidence of open-world classifiers (product identification, QC automation) and flag the uncertain predictions for human QC, across 100k+ images/day.
• Fine-tuned YOLO and the transformer-based RF-DETR with hard-negative mining and multi-GPU cloud training, to suppress confirmed empty-shelf false positives without degrading other shelf classes.
• Built the RGB+depth version of a classifier that checks empty-shelf alerts, with depth estimated by Depth Anything V2 as a fourth input channel, to tell true gaps from dark or shadowed shelves.
• Designed automated relabeling pipelines that update legacy bounding-box annotations at scale, adapting large datasets to newer architectures without manual re-annotation.

Conformal and adaptive soft sensors for real-time quality prediction in industrial batch processes. Supervised by Guillaume Ramelet and Dr. Maxime Moreaud.

• Introduced Self-Organized Conformal Prediction (NeurIPS 2026): unsupervised group discovery with a Self-Organizing Map for group-local calibration, cutting regional coverage gaps on all 10 datasets, by 14.8% on average, for only 3% larger prediction sets.
• Co-developed Mask-Valid Conformal Prediction (with M. Hegazy, M. I. Jordan, A. Dieuleveut): binary monitoring masks that interpolate between marginal and anytime coverage, with restart validity after failures and no need for past test labels.
• Created and open-sourced TorchSOM (JMLR 2026), a PyTorch library for Self-Organizing Maps: 81-98% faster than MiniSom on CPU, up to about 12× faster than somoclu on GPU, and the lowest topographic error in every rectangular-map benchmark (62-100% below MiniSom).
• Developed adaptive Just-in-Time soft sensors pairing SOM similarity retrieval with online SHAP feature selection, cutting MSE by up to 61% over global baselines (IEEE ETFA 2026) and prediction error by over 90% under year-long drift on 35k production batches.
• Built an explainable analytics framework (frequency-ranked feature selection, SHAP interactions, multi-split conformal prediction) on 35k rubber-mixing records, cutting 316 input variables to 20 at equal accuracy (Digit. Chem. Eng. 2026).

Led the lab sessions of Introduction to Machine Learning and Data Science for first-year engineering students (B3/MSc1): 20h a year for three consecutive years, 20 to 25 students per session. Supervised by Pr. Philippe Xu.

• Taught the fundamentals of machine learning, from data preprocessing and feature engineering to model training and evaluation.
• Guided students through their first hands-on experience building and validating predictive models.

Model-based uncertainty quantification for sample-efficient quality-diversity optimization in robotics. Supervised by Pr. Antoine Cully.

• Tackled the data-hunger of MAP-Elites by integrating probabilistic surrogate models on a robotic arm control task; the best variant beat MAP-Elites and came close to MAP-Elites with sampling.
• Benchmarked eight algorithmic variants combining different surrogate architectures and acquisition strategies in non-deterministic environments.
• Introduced two metrics for the uncertainty of quality-diversity archives.

Deep learning for detecting fast ripples, pathological biomarkers for epileptic tissue, in intracranial EEG signals. Supervised by Dr. Ludovic Gardy, Pr. Christophe Hurter, and Pr. Emmanuel Barbeau.

• Trained CNN fast-ripple detectors on wavelet time-frequency images of macro- and micro-electrode EEG, improving accuracy by 5% through systematic hyperparameter tuning (J. Neurosci. Methods 2025).
• Applied Grad-CAM to localize high-frequency bursts in time-frequency maps, giving clinicians visual explanations anchored to known fast-ripple signatures.
• Accelerated EEG signal and image processing by 15% with Cython, and built 7 NoSQL databases for multi-channel EEG at macro and micro electrode scales for batch retrieval and parallel processing.

Anisotropic image filtering for sub-pixel accurate line feature detection, published at IEEE ICASSP 2023. Supervised by Pr. Baptiste Magnier and Pr. Binbin Xu.

• Designed the 2D Semi Bi-Gaussian (2DSBG) filter, an asymmetric convolution kernel that enhances elongated structures while suppressing isotropic noise.
• Benchmarked against state-of-the-art edge detection models (DexiNed, RCF, Laplacian-of-Gaussian) on synthetic and real-world datasets.

Applied Mathematics, Machine Learning, Research

• Industrial PhD (CIFRE) with Michelin: adaptive and conformal soft sensors for real-time quality prediction in multistage industrial processes.
• Under the supervision of Pr. Eric Moulines and Dr. Ahmed Shokry.
• Adaptive Soft Sensing, Conformal Prediction, Explainability, Self-Organizing Maps, Online Learning, Dimensionality Reduction

Computer Science, Mathematics, Research

• Diploma with merit (2:1), and thesis with distinction (1st).
• Research internship at the Adaptive & Intelligent Robotics Lab (AIRL), under the supervision of Pr. Antoine Cully.
• Machine & Deep Learning, Reinforcement Learning, Computer Vision, Natural Language Processing, Mathematics for Machine Learning, Computational Finance, Robot Learning.

Mathematics, Data Science, Research

• First-class honours with jury's congratulations
• Ranked 2nd of 60 in the CS & AI department, top 5% of the school
• Research internship at the CNRS (CerCo), under the supervision of Dr. Ludovic Gardy, Pr. Christophe Hurter and Pr. Emmanuel Barbeau.
• Advanced Probability & Statistics, Operations Research, Algorithms & Complexity, Machine & Deep Learning, Software Engineering

Selected for the inaugural cohort of this year-long international research fellowship. Project: generative and world models for EEG, with fellow researchers and Paul Hellegouarch, mentored by Dr. Quentin Garrido and Dr. Jean-Baptiste Masson.

Built a hierarchical JEPA world model on the complete fly-brain connectome with team Piaget; reposted and congratulated by Yann LeCun.

Selected from 1,700+ applicants; built Cerno, an AI recruitment platform with adaptive voice-AI interviews, in 36 hours.

Won the audience vote for best 3-minute PhD presentation at Institut Polytechnique de Paris, selected to represent IP Paris in the national round.

Best poster in the Mathematics category at the IP Paris Welcome Day, for work on explainable quality prediction with uncertainty quantification.

First place at Michelin’s internal AI hackathon for manufacturing innovation.

Scholarship supporting international academic mobility.

Merit-based scholarship from the Mines Alès alumni association.

Academic excellence scholarship.

Creativity laureate for a Communication & Media competition.

• Led a team of 10 photographers covering 25 university events.
• Edited and produced videos and 1,000+ photos.

• Led a 25-person team running 5 tournaments across 20 sports
• Secured partnerships with brands, including Armor Lux and Tony Parker.
• Managed the website and IT infrastructure.

• Raised over €3,500 for solidarity projects in Madagascar.
• Organized and participated in fundraising events, including cooking for 300 students.

• Achieved 3rd place in the national university league in the UK (BUCS).
• Achieved 4th place in the national university league in France (FFSU).
• Won the 49th Challenge Centrale Lyon
• Won the 19th TRAMS

• Prepared event logistics and setup for a multi-school rugby tournament.

• Provided music and technical support for student events and gatherings.

• Twice finalist in a male modelling contest.

Co-supervised my work on computer vision for open-world retail deployments.

Email: name at augmodo dot com

Co-supervised my PhD thesis on conformal and adaptive soft sensors for real-time quality prediction in multistage industrial processes.

Email: name dot surname at polytechnique dot edu

Co-supervised my PhD thesis on conformal and adaptive soft sensors for real-time quality prediction in multistage industrial processes.

Email: name dot surname at polytechnique dot edu

Co-supervised the start of my PhD thesis on conformal and adaptive soft sensors for real-time quality prediction in multistage industrial processes.

Email: name dot surname at michelin dot com

Co-supervised my PhD thesis on conformal and adaptive soft sensors for real-time quality prediction in multistage industrial processes.

Email: name dot surname at michelin dot com

Supervised my independent research project and my MSc research internship on quality-diversity optimization.

Email: a dot surname at imperial dot ac dot uk

Co-supervised my MSc research internship on ML-based pathological oscillations detection in epilepsy.

Email: name dot g at sotisai dot com