Paris, France · Open to chat

building RecSys@Criteo

just an engineer who likes agents, fast code and GPUs going brrr

Get in touch
8 stages

What I build

Two loops.Offline, a sequential model is retrained every day on fresh session sequences and rolled out progressively; when a new configuration beats it on offline metrics, it goes to an A/B test.Online, your last actions collapse into one embedding, and an approximate nearest-neighbor search returns the closest products before the page finishes loading.

Offlinetraining
Event logs
views · clicks · carts
Data preparation
→
raw events → session sequences
Training
step 0k · converging
Offline eval
recall@500.000
nDCG@100.000
AUC0.000
held-out sessions
Deploydaily rollout
inference server · both versions loadedrequests target v1
v1yesterday · serving
modelann index
v2today
modelann index
qps · last 30 minv1v2
rollout · both serving
−30 minnow
Onlineserving
User sequence
t−4
t−3
t−2
t−1
t
last actions, in order
Inference
→
N item vectors → 1 user vector
ANN search
probing 2/4 cells · 50% scanned
hitmissedfalse hitdashed = true top 3
A/B test
ctrl
50%
var
50%
ctr uplift · 80% ci+0.0%
50/50 · new configsignificant
Costtune it yourself
recall@367%2/3 true neighbors found1 sits in a cell we didn’t probe
serving cost×1.00vs baseline
latency budget25% used
banner servedpersonalised
the nearest neighbors made it back in time
ann 61%ffn 26%attention 8%fetch 5%
  1. 01 Event logsRaw user actions, partitioned by day
  2. 02 Data preparationCleaned and grouped into session sequences
  3. 03 TrainingLoss converging on held-out sessions
  4. 04 Offline evalrecall@k · nDCG · AUC
  5. 05 User sequenceThe last actions in the session
  6. 06 InferenceOne pass → one user embedding
  7. 07 ANN searchApproximate nearest neighbors in the index
  8. 08 A/B testNew config wins offline → 50/50 on live traffic

Where I work

2024
Criteo logo

Criteo

Apr. 2024 - Current · Paris, France

Senior Machine Learning Engineer

Oct. 2026 - Current
  • Reco Retrieval team (AI for Performance group), owning retrieval models end to end, from research idea to worldwide production
  • Lead the design of new retrieval systems and write the design docs that align research, engineering and product
  • Build sequential user models that learn from shoppers' latest actions to retrieve relevant products in real time
  • Run large-scale training pipelines (data preparation, daily retraining on fresh data, offline evaluation) and keep making them faster and cheaper
  • Launch A/B tests up to worldwide traffic, analyze them statistically, and turn the results into ship decisions
  • Deploy models to production and optimize online inference to fit strict latency constraints
  • Monitor model quality and serving health across offline and online systems, and act as on-call engineer when production breaks
  • Mentor engineers and researchers, and write the documentation the team relies on
  • Ship fast and iterate faster by putting AI agents into every step above

Machine Learning Engineer II

Oct. 2025 - Oct. 2026

Machine Learning Engineer

Oct. 2024 - Oct. 2025

Machine Learning Research Intern

Apr. 2024 - Oct. 2024
4 institutions

Where I studied

2023
Ecole Normale Supérieure Paris‑Saclay logo

Ecole Normale Supérieure Paris‑Saclay

Sep. 2023 - Sep. 2024 · Paris, France

MSc in Mathematics, Vision, Apprentissage (MVA)

  • World's leading #1 university in Mathematics (ARWU)
  • Courses: Deep Learning (theory and practice), Computer Vision, Object Recognition, Reinforcement Learning, Time Series Analysis, Natural Language Processing, Speech Processing, Convex Optimization, Probabilistic Graphical Models, Deep Generative Models, Signal Processing
  • Representative of a class of 250+ students
Interviewed in Le Monde
2018
INSA Centre Val de Loire logo

INSA Centre Val de Loire

Sep. 2018 - Sep. 2023 · Bourges, France

MEng in Industrial Systems

  • Courses: Statistical Theory and Methods, Machine Learning, Computer Vision, Complex Systems Engineering, Risk Management
  • Student representative, active participation in various events (student council campaign, open days, etc.)

During this programme

2022
University of Tours logo
University of Tours

MSc in Applied Physics (Double Degree)

Sep. 2022 - Sep. 2023 · Tours, France

  • Courses: Automatics, Robotics, Cyber‑Physical Systems, Multiphysics Modelling
2021
Tallinn University of Technology logo
Tallinn University of Technology

School of Information Technologies (Erasmus)

Sep. 2021 - Jan. 2022 · Tallinn, Estonia

  • Courses: Data Mining, Machine Learning, System Programming, Embedded Systems