- Lead causal measurement of the productivity impact and ROI of AI tools — coding assistants and internal agents — on topline metrics across Meta's family of apps.
- Build quasi-experimental and probabilistic models — difference-in-differences, synthetic control, interrupted time series, state space models, Gaussian processes, IRT models — to quantify AI-driven impact at scale.
- Develop interpretable production forecasts and agentic decision support whose outputs are inputs to exec-level AI investment decisions.
Alexandre Andorra
Machine Learning Scientist
Probabilistic ML, Causal Inference & Model Evaluation
andorra.alexandre@gmail.com LinkedIn GitHub alexandorra.github.io
Machine learning scientist with 8+ years building and evaluating production probabilistic AI systems across tech, pharma and professional sports. I specialize in Bayesian ML, causal inference, time series and rigorous model evaluation. At Meta, I measure the causal impact of AI tools on topline metrics across the family of apps. Core contributor to PyMC (8K+ GitHub stars), co-founder of PyMC Labs ($0 → $1M annual revenue in 2 years), principal modeler on a paper under review in Nature, and creator of Learning Bayesian Statistics, a top-1.5% global podcast.
Experience
- Built production Bayesian forecasting models on 1M+ tracking events (Trackman, HawkEye) to predict player performance, improving projection accuracy by 15% over baseline.
- Developed hierarchical time series models integrating player tracking, biomechanics and game data to inform $100M+ roster decisions.
- Translated probabilistic model outputs into actionable insights for coaching staff and front office, directly influencing game strategy and player development.
- Led technical education initiatives upskilling the 10-person R&D team in advanced statistical methods.
- Co-founded and scaled the company from $0 to $1M annual revenue in 2 years (50% YoY client growth), securing 15+ enterprise contracts and establishing it as the leading Bayesian ML consultancy.
- Architected production systems: Bayesian A/B testing frameworks, causal inference pipelines (treatment effects, substitution effects, backdoor adjustment), Gaussian process models for spatio-temporal forecasting.
- Delivered for Fortune 500 clients in pharma R&D (RNA vaccine development), agtech (crop optimization) and e-commerce (marketing attribution reaching 10M+ users), leading cross-functional teams.
- Designed and delivered 40+ technical workshops training 500+ practitioners in hierarchical modeling, causal inference, time series and decision optimization.
Earlier: Data Scientist at Société.com (2020–21), 360Learning (2019–20) and Banque de France (2018–19): financial prediction models, ETL pipelines, automated financial reporting.
Research & Open Source
Open-source agent skills that teach AI coding assistants rigorous Bayesian modeling, causal inference and model evaluation; works with Claude Code, Codex, Cursor and Gemini CLI.
Co-author and principal modeler (bespoke Bayesian hierarchical model); under review in Nature.
Hierarchical Bayesian model separating player skill from team effects over time; introduced Skill and Performance Above Replacement; interactive rankings of 1,000+ players with uncertainty.
Implemented the ZeroSumNormal distribution, improved the multivariate normal API and created prior helper functions. Authored 8 official tutorials (HSGP, multilevel modeling, Kronecker covariances, LKJ priors) used by thousands of practitioners.
First French polling aggregator, built on MRP models, with GP and HMM models of presidential approval; fully open source; 4K visitors/month during campaigns.
Teaching & Community
Top-1.5% global podcast and the largest Bayesian statistics community: 12K monthly listeners, 165+ episodes, 100+ paid community members. Currently hosting a 12-episode series on reinforcement learning.
Online platform teaching probabilistic AI to 200+ students ($100K+ revenue).
Selected Talks
- Was It the Player or the Team? Building a Bayesian Skill Model for Soccer — Carnegie Mellon University, live workshop, Oct 2026
- What Your Mixed Model Already Believes — Visiting Speaker Seminar, Princeton University, Sep 2026
- Amortized Inference & GPUs for Stan — live at StanCon, Aug 2026
- A Beginner's Guide to State Space Modeling — PyData Berlin, 2025
- Mastering Gaussian Processes with PyMC — PyData NYC, 2024
- Live podcast show — StanCon Oxford, 2024
- Guest on Stats + Stories and the Super Data Science podcast
Skills
- Model evaluation
- Proper scoring rules, calibration (PIT, LOO-PIT), LOO-CV and model comparison, posterior predictive checks, A/B testing and experimentation, sequential analysis
- Causal inference
- Difference-in-differences, synthetic control, interrupted time series, propensity scoring, backdoor adjustment, DAGs
- Machine learning
- Bayesian ML, probabilistic AI, hierarchical models, Gaussian processes, state space models, IRT, Bayesian optimization, deep neural networks, AI agents
- Tools
- Python, PyMC, NumPyro, JAX, PyTorch, TensorFlow, NumPy, pandas, scikit-learn