Sequential prediction models
Next-generation conversion-rate models and large-scale user representations for real-time bidding.
Production machine learning
I build ML systems that ship.
Machine Learning Engineer at Criteo, working across deep model architecture, large-scale training, experimentation, and low-latency serving. I turn research ideas into measurable product impact.
Project ownership
Systems I have led end to end.
From problem framing and architecture to experimentation, rollout, and long-term iteration.
Deep bidding architecture
Multi-head and multi-phase model designs, including migration paths for high-throughput production bidders.
Training and evaluation platforms
Data pipelines, offline evaluation frameworks, and training optimization balancing model quality, speed, and cost.
Cross-team model strategy
Technical roadmaps, production rollout strategy, performance reviews, and research mentorship across multiple teams.
Featured publication · WSDM 2026
Abacus: Self-Supervised Event Counting-Aligned Distributional Pretraining
A counting-aligned pretraining objective for sequential user models, co-authored at Criteo and evaluated on real-world advertising datasets.
Experience
Research depth, production scale.
From unsupervised representation learning to real-time advertising systems.
Nov 2023 — present
Machine Learning Engineer
Criteo · Paris
Design and optimize production-scale deep bidding, attribution, and sequential user-modeling systems—from model architecture and data pipelines through experimentation, rollout strategy, and cost-efficient inference.
Jun 2024 — present
Expert Member · Earth Science Review Board
NASA Lifelines · Volunteer
Support humanitarian organizations as an Earth Observation expert, including satellite-imagery prototypes for road accessibility and technical reviews.
Oct 2020 — Oct 2023
PhD Researcher in Deep Learning & Satellite Imagery
CentraleSupélec / ONERA · Paris
Developed unsupervised representation-learning methods for Sentinel-1 SAR time series, from agricultural monitoring to boreal forests, including a visiting research period at ESA Φ-lab.
Technical focus
Deep learning, end to end.
Architecture and training are only useful when they survive production constraints.
Modeling
PyTorch
Sequential Models
Representation Learning
pCR / CVR
Self-Supervised Learning
Data & Experimentation
Python
SQL
A/B Testing
Evaluation Design
Distributed Computing
Ray
Distributed Data Parallel
PySpark
Production
Low-Latency Inference
Training Optimization
Kubernetes
Cost Optimization
Open source
Tools beyond production work.
A focused utility that makes multitemporal SAR analysis easier to reproduce.
Python
Google Earth Engine
Sentinel-1
GEE SAR Fetcher
A Python library for downloading preprocessed Sentinel-1 temporal stacks directly from Google Earth Engine, removing repetitive calibration and coregistration work from SAR time-series analysis.
pip install geesarfetcher
Sentinel-1
→
Google Earth Engine
→
Python stacks
Get in touch
Let's talk ML systems.
For engineering, applied research, or collaboration conversations, email and LinkedIn are the best ways to reach me.