Portfolio
Production AI systems, data science, and compliance
Selected work
RAG System
Three retrieval strategies compared over one corpus: classic, graph and agentic. Evaluation harness with a golden dataset, LLM-as-judge scoring and regression detection, shipped with GitHub Actions CI running ruff, mypy and pytest.
Automated Application Catalogue (Kabeen)
Professional work during my alternance at Kabeen, also defended as my Epitech end-of-studies project. Hybrid-search RAG/ETL pipeline for metadata extraction and classification over a 100,000+ entry application catalogue, with scraping-based ingestion, data-quality and anomaly detection, and LLM modules in production. 80% reduction in document processing time.
NER Travel Order
CamemBERT fine-tuning for French named entity recognition. BIO tagging, token classification over a 12,000-sentence annotated corpus. spaCy F1 0.84 (~5 ms) vs CamemBERT F1 0.89 (~50 ms); spaCy selected for production on the latency/size trade-off.
E-Invoicing Factur-X
French e-invoicing compliance (QR PDP standards). LLM-assisted data capture, multi-role agent exploration, production pipeline.
More work
On-Device Moderation
On-device image moderation for a messaging app. MobileNetV3Small (frozen backbone + trained head), 3 classes over 9,714 images. 93.1% macro F1, exported to TFLite in three quantised variants (1.24 / 1.96 / 3.89 MB) for edge, mobile and server.
Chest X-Ray Classifier
Deep learning CNN for medical imaging classification. Data augmentation and transfer learning.
Sentiment Analysis BERT
Fine-tuned BERT transformer for sentiment classification. Built with PyTorch and Hugging Face.