Julie B.

Internship within the Toxicology and Data Science teams. I transformed an internal Excel tool, used to estimate the skin penetration of cosmetic molecules from a physico-chemical equation, into a performant and maintainable Python API. This work involved complete code refactoring, automation of calculations, implementation of unit tests, and performance optimization. I also ensured the integration of the API into the company's environment and wrote clear technical and scientific documentation for developers and end-users.

I built a complete pipeline to classify land cover (urban, agricultural, natural) from Sentinel-2 multispectral images: extraction of spectral indices (NDVI, SAVI) and edge filters (Sobel), manual zoning on QGIS, raster → feature table conversion, then training/optimization of SVM and Random Forest models. The SVM model achieved over 85% accuracy, with interpretable final mapping and documentation ensuring reproducibility.

Project aimed at automating the identification of four types of white blood cells (eosinophils, lymphocytes, monocytes, neutrophils). From a Kaggle dataset of 12,500 annotated images, I: (1) preprocessed the images (resizing, normalization, augmentation, and class rebalancing); (2) designed a convolutional neural network using Keras/TensorFlow achieving 96.5% accuracy on the test set; (3) generated confusion matrices, F1-scores, and visualizations to analyze confusions; (4) compared the CNN to simpler Scikit-learn models, demonstrating substantial gains; (5) documented the pipeline and proposed avenues for reducing the overfitting observed on certain classes.