AI/ML Engineer in Paris · LLM agents, generative models and evaluation · M.Sc. AI, Université Paris-Saclay
- Text-to-SQL agent – Capital Fund Management (2025). LangGraph agent with 7 tools that answers analysts' questions over an internal Oracle database: 87.5% accuracy on a 48-question benchmark I built, about 20× cheaper per query than o3 at similar accuracy. Code is internal; internship report (PDF).
- First text-to-speech system for Walloon – LISN, CNRS (2024). VITS trained on 18 and 156 minutes of speech; fine-tuned from French, 4.10/5 MOS with 23 listeners (natural speech: 4.48–4.55). First-author paper, SSW 2025. Listen · Code · Paper · Demo
- AskData (in progress). Ask a PostgreSQL database questions in plain English: schema retrieval, planner-guided SQL, read-only validation. FastAPI, Next.js. Repo
- Synthetic French NER dataset for clinical trials (2024, team of 3). Translated English annotations, fine-tuned XLM-RoBERTa, released a 12,423-row dataset. Repo
Now: research collaborator at Universidad Tecnológica de Bolívar, lead author of a paper on cross-lingual embeddings (under review).
Stack: Python · PyTorch · Hugging Face Transformers · LangGraph · FastAPI · PostgreSQL · Docker · Git · Linux
Website · LinkedIn · Google Scholar · Hugging Face · felipeespinosa.13@outlook.com
