Graduate engineer with a dual MSc from École Centrale Marseille and DTU, specialized in AI, data science, and computer vision with a strong background in applied mathematics and physics. Experienced in developing end-to-end ML pipelines, image and signal analysis, and time-series modeling in both research and industrial environments. Key projects include spectroscopy-based material classification (Richemont), hyperspectral imaging for agrifood and life sciences (Videometer), and biomedical image segmentation (Inria). Skilled in Python, PyTorch, TensorFlow, HPC, and statistical modeling.
A selection of projects from my academic, professional, and research experience.
Research project at Richemont. Developed AI models to identify and characterize materials such as metals and gemstones from LIBS spectroscopic data. Built end-to-end pipelines—from data preprocessing to evaluation—using machine learning techniques such as classification, clustering, and anomaly detection. Emphasis on chemometric analysis. Tools: Python (PyTorch, Scikit-learn) and GitLab.
Academic project in collaboration with Videometer, a leader in hyperspectral imaging. This project focused on detecting and classifying healthy vs. unhealthy maize kernels in multi-spectral conveyor belt images using deep learning. It combined selective search with a VGG-16-based CNN, followed by non-maximum suppression for refined predictions. The work demonstrated the feasibility of deep learning for precision quality control in agriculture.
Research internship at Inria. Worked on the refinement and evaluation of a cell segmentation and tracking pipeline for 4D organoid images acquired via light-sheet microscopy, using watershed and morphological operations. The image on the left shows the temporal evolution of a 3D segmentation of an organoid. The project is referenced in the 2022 Inria Annual Report.
Academic project focused on binary segmentation of retinal blood vessels using the U-Net architecture. A dataset of 20 labeled retinal images was preprocessed with CLAHE (Contrast Limited Adaptive Histogram Equalization) and normalized before training. The model was trained for 40 epochs using a binary cross-entropy loss and Adam optimizer. Evaluation on the test set yielded a Dice coefficient of 0.78, highlighting U-Net’s effectiveness for vascular segmentation in biomedical imaging.
Academic project analyzing urban expansion in Dubai’s Palm Islands using multi-temporal satellite imagery accessed via Google Earth Engine. The methodology involved applying Canonical Correlation Analysis (CCA) to detect and quantify land cover changes between 2000 and 2008. The resulting change map effectively highlights spatial patterns of artificial land growth and urban development.