Portfolio
What I do, and what it has produced so far.
Every number on this page comes from a controlled experiment or from what the company reported for that role. The orange ones are measured treatment effects, the rest is scale. If you want the long version of any of it, the case studies are below, and I'm happy to walk through them on a call.
Six things I'm good at, each with a result behind it.
- 01
Experimentation and A/B testing
Designing tests that can actually answer the question: the right metric, enough traffic, sensible guardrails, and a clear call at the end. Mostly on booking and checkout flows with tens of millions of visits a month.
+0.6% conversion, +0.5% revenue per visitor (relative) · Air France-KLM, one-month A/B tests on hundreds of thousands of visitors a day
- 02
Causal inference and uplift modelling
Meta-learners, heterogeneous treatment effects, persuadables. The point is to spend the budget on people whose behaviour will change, not on people who would have bought anyway.
+30% relative conversion uplift · Senzai, telecom client; +5–15% monthly for a bank · published at ICML 2023
- 03
Recommender systems and personalisation
Ranking and offer recommendation on booking, checkout and CRM flows, evaluated offline first and then against a control group before anyone rolls it out to everyone.
€1M+ incremental revenue per month · Air France-KLM flight ranker and branded-fare recommender
- 04
MLOps and production ML
Airflow, MLflow, DVC, Docker, Kubernetes and GitLab CI/CD, on AWS, GCP or Azure. Pipelines that score millions of customers and let a new model ship in days rather than weeks.
3M customers and 50M interactions scored in production · Senzai, ML lead
- 05
Bayesian modelling and uncertainty quantification
Gaussian processes, calibrated prediction intervals and sensitivity analysis, for the decisions where a point estimate isn't enough and someone needs to know how sure we are.
80% prediction intervals for well production · TotalEnergies · journal paper in Computational Statistics and Data Analysis, patent application US 2024/0211803 A1
- 06
Large-scale data engineering
Feature engineering and pipeline optimisation on tens of millions of rows, with BigQuery, PySpark, Redshift and PostgreSQL.
Detection runtime cut from 25 h to 8 h · namR, a full-factorial plan replaced by an optimised Latin-hypercube design on 50M-row datasets
The projects in detail.
Personalising flight and fare offers with A/B-tested recommenders
Two recommender systems on the booking flow: which flights to show first, and which fare to suggest. I work on the metrics and segments, prepare the data, and run the offline and online evaluation of every A/B test before anything ships.
Uplift-based targeting for digital marketing campaigns
I led the machine learning of a re-engagement engine: the recommendation system that decides which customers to contact, when, and through which channel, built on causal inference rather than propensity, plus the pipelines, evaluation and CI/CD it runs on.
Causal inference and uncertainty quantification for energy production
Three years of research where statistical learning meets causal inference: calibrated prediction intervals for gas-well production, and treatment-effect estimation for geothermal wells. It led to a patent application and papers at ICML and in CSDA.
Evaluating frontier language models on mathematics, agents and data
Three projects for AI research labs: writing mathematics problems that frontier models get wrong and experts can solve, evaluating models and agentic frameworks under constraints, and a STEM review project producing data visualisations with the science checked behind each one.
Methods, languages and infrastructure.
- Methods
- Languages
- Libraries
- MLOps and production
- Cloud and data
- Tools
Sound like your problem?
I take on freelance and consulting work in marketing analytics, experimentation and causal ML, anything from a single A/B test to a full personalisation programme. Project-based or longer, remote or in Paris, France. Send me an email and we'll set up a short call.