projects
completed projects
RegiBERT: Continuous Modeling of French Linguistic Registers via Label Distribution Learning, Linear Probing and UMAP Projection
The detection of linguistic registers has historically suffered from rigid classification approaches (hard labels). However, natural language is inherently nuanced and ambiguous. This project introduces RegiBERT, a CamemBERT-based model designed to capture this linguistic continuity. Rather than assigning a single exclusive class to a sentence, the model learned to predict a probability distribution (soft labels) using the annotated TrémoLo corpus. RegiBERT achieves a validation KL divergence loss of 0.365, a MAE of 0.179, and an overall accuracy of 84.5%. Layer-wise probing further reveals that stylistic nuance crystallizes in deeper transformer representations, peaking at layer 12 (R² = 0.556). A 3D UMAP projection calibrated on the validation subdataset confirms this structure geometrically, revealing a continuous manifold where the standard register smoothly connects the formal and colloquial clusters.
Empirical Benchmark: Johnson-Lindenstrauss Lemma on Text Data
Practical benchmark of dimensionality reduction techniques on text data (20newsgroups). Compares PCA against Random Projections, proving that Random Projections preserve pairwise distances while running much faster than standard SVD in high dimensions.
L'Exil Scolaire : Overseas Students' Journey (INSEE Dataviz 2025)
Finalist project (3rd / 17) for the national INSEE Dataviz Competition. Conducted an in-depth statistical investigation into educational mobility inequalities and structural gaps facing students moving from French Overseas Territories to Mainland France. Developed custom visual flows and spatial representations to highlight regional disparities in higher education access.
ongoing projects
MLP, CNN and Transformer implementation from scratch
The goal of this project is to learn the fundamental architectures of deep learning (MLP, CNN, Transformer) by an implementation entirely in NumPy (forward and backward passes coded and derived by hand), without frameworks such as PyTorch or TensorFlow. My studies will mostly be based on Standford's cs231n notes.
FairCut : Autonomous Algorithmic Fairness & MLOps Pipeline
Production-ready MLOps infrastructure predicting daily box-office success via TMDB and XGBoost while auditing algorithmic bias across demographic cohorts using Fairlearn and Evidently AI. Features automated daily ingestion, data drift detection, continuous model retraining, automated promotion via MLflow, and full IAC deployment with Terraform and GitHub Actions onto an interactive Streamlit dashboard. (Collab: Lina Lebbah)