松涛 林

Development of an ESG Data Extraction Tool (API server) for ESG Data (LLM, RAG, extraction API, structuring of financial reports): - Created from scratch a pipeline (based on DVC) to standardize the processes of retrieval, extraction, and training of reranking models, featuring a well-organized visual interface. - Reduced hallucination rates by training an XGBoost model as a Reranker, combining it with a dynamic threshold method to reject false positives. Improved retrieval accuracy by 21% and extraction accuracy by 6%. - Explored and implemented new solutions: precise extraction of tabular structures (using a YOLOv10 model, pdfplumber library) and prompt engineering to overcome challenges in extracting cross-page tables and multi-year data. - Integrated and deployed in production: CI/CD pipeline (GitLab, Docker, Kubernetes) deployed in the operational environment with rigorous version control.

Réalisation d'un site web connecté à une base de données archéologique en utilisant PHP pour le serveur, et jQuery, HTML, CSS pour le front-end. Construire et maintenir un entrepôt de données selon les exigences, en formant un système de données surveillable, maintenable et actualisable.