Engineering student at CentraleSupélec (Université Paris-Saclay), previously Sorbonne Université. I build applied AI and real-time systems that have to work outside the notebook — for uranium mines, financial advisors, and a live-streamed ATM.Engineering student at CentraleSupélec. I build applied AI and real-time systems that have to work outside the notebook — for uranium mines, financial advisors, and a live-streamed ATM.
Looking forRole
6-month role in applied AI / machine learning — internship, VIE or fixed-term contract6 months in applied AI / ML — internship, VIE or fixed-term
A 24/7 live-streamed sculpture fed by an internet-connected ATM. I built it end to end, including a low-cost streaming stack whose cost stays flat regardless of audience size.
Deep learning for real operationsDeep learningOrano, 2026
A physics-informed model for uranium mine operations
Asked to fix a production model planning well cleaning, I traced its errors to interactions between neighboring wells — then derived a new model from the geology literature.
Transformer with spatio-temporal cross-attention + Neural ODE
Data
10,000+ well time series, up to 20 years each10,000+ well series, up to 20 years
Found
~30% of legacy data silently imputed
02 — About
Math, applied AI, and systems that ship.
I started programming at 12 and never really stopped. I did a double bachelor's in mathematics and computer science at Sorbonne Université, then joined CentraleSupélec (Université Paris-Saclay), where I'm on a gap year before my final engineering year. The mathematics is what I lean on most — usually when an AI model does something I wasn't expecting.
Right now I'm a data science intern at Orano and lead developer at QSTNMRK, on art e-commerce and live-streaming projects. A few years of freelancing before that taught me to pin down what a client really needs, and to stay on the hook when something breaks after delivery.
Next, I want to work close to the people who have the problem — on hard technical problems, across different clients. I'm looking for a 6-month role from February or March 2027 — internship, VIE or fixed-term contract all work for me — in Asia, the US or Canada.Next, I want to work close to the people who have the problem — on hard technical problems, across different clients.
Global nuclear fuel cycle company, from uranium mining to recycling.orano.group ↗︎
Traced the errors of a production model planning well cleaning across uranium mines to an overlooked factor: interactions between neighboring wells.
Designed a physics-informed replacement — a Transformer with spatio-temporal cross-attention coupled to a Neural ODE, trained on 10,000+ well time series spanning up to 20 years.
Built the extraction from mine monitoring systems (SQL, REST) and found ~30% of legacy data silently imputed.
PyTorch
Transformers
Neural ODE
Physics-informed ML
SQL
Claude Code + MCP
Jan 2024 — Present
Part-time
Lead Developer · QSTNMRK
Art collective — e-commerce, IoT and live-streamed works.
Built The Last Dollar end to end, featured in designboom: ATM hardware over MQTT, and a streaming stack whose cost stays flat regardless of audience size.
A small delivery vehicle on Raspberry Pi and Arduino that detects its track by image processing, computes steering, and drives the motors through turns in an urban-like course.
Computer vision
Raspberry Pi
Arduino
Control
Jun 2024Private
Graphical Models for Financial Dependency
Estimates conditional dependencies between MSCI World assets with Graphical Lasso. The recovered network hinges on regularization, so most of the work was model selection and rolling-window validation.
Python
Graphical Lasso
Statistics
Finance
May 2024 · with Bouygues TelecomMay 2024 · Bouygues TelecomPrivate
Anomaly Detection on a LoRaWAN IoT Network
A Spark pipeline over large-scale LoRaWAN logs that flags devices and traffic behaving abnormally and points at likely causes — unsupervised, on time-series features.
A recommender built on subtitle text rather than ratings, with TF-IDF, K-Means and a perceptron. Supervised by Nicolas Baskiotis; results turned out to hinge on preprocessing.