What I do
I help companies design and build machine learning and LLM-based systems, from scoping and prototyping to deployment. My background is in mathematical statistics: PhD on survival analysis (LPSM, supervised by Olivier Lopez) and agrégation de mathématiques.
I currently work on:
- LLM agents and retrieval systems: agentic workflows, hybrid search, human-in-the-loop design, evaluation, sovereign deployments with open-weight models.
- Predictive modelling: forecasting, predictive maintenance, risk scoring, including messy real-world data.
- Statistical modelling: survival and time-to-event analysis, churn, censored data, uncertainty quantification.
Experience
- Independent consultant (2025–present). LLM-based agents for compliance workflows in the financial sector, with human-in-the-loop review and sovereign deployment.
- Senior Data Scientist, C3 AI (2022–2024). Predictive maintenance for a global cement producer: per-asset models on 200+ assets with a custom MLOps layer, and a 90% reduction in false-positive alerts. Supply-chain risk forecasting, with a 40% reduction in MAE over the baseline.
- Machine Learning Researcher, Zelros (2019–2022). AI for insurance: automated 40% of medical underwriting decisions, built a semantic search engine and a bandit-based recommender, and created CinnaMon (see below).
Research interests
Foundation models for tabular and time-to-event data, AI agents, evaluation of LLM agents, and quantitative finance (market microstructure, systematic strategies).
Publications
- Journal: G. Gerber, Y. Le Faou, O. Lopez, M. Trupin (2020). The impact of churn on client value in health insurance, evaluation using a random forest under random censoring. Journal of the American Statistical Association. [code] • [talk – IME 2017]
- Preprint: Y. Le Faou, O. Lopez (2018). A nonparametric conditional copula model for successive duration times, with application to insurance subscription. [code] • [poster – DS3 2018]
- PhD thesis: Y. Le Faou (2019). Contributions à la modélisation des données de durée en présence de censure : application à l'étude des résiliations de contrats d'assurance santé. (Contributions to the modelling of censored duration data, with an application to health insurance contract terminations)
Open source
- Décrypter l'Assemblée (2026): open-source tool to explore voting data from the French National Assembly. [site] • [GitHub]
- CinnaMon (2021): Python library for monitoring ML systems, focused on data drift. Presented at the Paris Machine Learning Meetup (2022). [GitHub] • [video, in French] • [slides]
- sword (2018): R package for regression on right-censored data. [GitHub]