Hi, I'm Killian!

A deep learning research engineer who loves Computer Vision :)

Learn about my journey in Mathematics, Computer Vision and Deep Learning

About me

About Me

I'm currently working on end-to-end multi object tracking at Idemia as an intern, to complete my master's degree in Machine Learning and Computer Vision at École Normale Supérieure Paris-Saclay (Master MVA).

I am passionate about computer vision and deep learning, driven by the conviction that visual perception is fundamental to intelligence, both human and artificial.

I strongly support open-source practices, transparency, and reproducibility in research. Sharing my work openly to allow others to build upon it is central to my approach to science.

Professional Experiences

Deep Learning Research Engineer (intern)
Idemia

April 2025 – October 2025

Enhancing end-to-end multi object tracking models by exploring various uses of SAM 2.

Automatic video instance segmentation mask generator: Use any MOT dataset and SAM 2 to automatically create high quality video instance segmentation masks for each tracklet.
Used such augmented datasets to train an end-to-end tracker (MOTIP) with an additional segmentation head, exploring various architecture.
Reproduced SAM2MOT, a zero-shot multi object tracker based on Grounding Dino and SAM2, to label any video dataset.
Explored ways to use SAM 2's embeddings to train a model based on MOTIP.

Intern – AI Research
CLS

April 2024 – August 2024

Foundation Models for remote sensing applications.

Developed a library to fine-tune Foundation Models on remote sensing datasets, used by coworkers at CLS.
Literature Review: Comprehensive review of latest research on Foundation Models in remote sensing.
Model Development: Developed a library for fine-tuning Foundation Models on remote sensing data.
Performance Evaluation: Evaluated fine-tuned Foundation Models for semantic segmentation tasks.

Intern – Machine Learning Engineer
JoliBrain

February 2023 – July 2023

Contribution to joliGEN: Implemented advanced edge detection and image generation controls from ControlNET.
Implemented state-of-the-art models such as Owl-ViT for zero-shot object detection.
Developed the documentation website for the tool using ReadTheDocs.

Intern – Software Developer (R Shiny)
French Ministry of Agriculture

May 2022 – August 2022

Created agreste: automation tool as an R package for streamlining statistical publication creation.
Managed project from requirements gathering to delivery, using Agile methodologies.
Designed an intuitive R Shiny UI to facilitate data input and publication generation.

Academic Background

Master 2 Mathématiques, Vision, Apprentissage
ENS Paris-Saclay

Sep 2024 – Mar 2025

Computational Optimal Transport, Convex Optimization, Turing Seminar (AI Safety)
Object Recognition and Computer Vision, Probabilistic Graphical Models, Deep Learning
Representation Learning, Generative Models, Data Generation by Transport/Denoising

Master 1 Applied Mathematics and Statistics
Toulouse School of Economics

Sep 2023 – Apr 2024

Econometrics, Probability and Statistics, Functional Analysis, Optimization for ML
Data Science (Python), Machine Learning, Mathematical Game Theory, Time Series

Gap Year
University of Copenhagen

Sep 2022 – Jan 2023

Courses: Natural Language Processing, Blockchain Business Development, Energy Economics, Tax Policy

Double Bachelor: Applied Mathematics and Economics
Toulouse School of Economics

Sep 2019 – Apr 2022

Linear Algebra, Analysis, Probability and Statistics, Optimization, Econometrics, Programming
Microeconomics, Macroeconomics, Numerical Analysis, Computer Science (Python, R)

Research Projects

View all research projects

Score-Based Generative Neural Networks for Large-Scale Optimal Transport

Max Daniels, Tyler Maunu, and Paul Hand, 2022

I performed several experiments to evaluate the performance of the SCONES method on low-dimensional distributions.

Test Time Training with Masked Autoencoders

Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei A Efros, 2022

I extended the TTT-MAE method with an online version that shows better performance than the original method.

Are Generative Classifiers More Robust to Adversarial Attacks?

Yingzhen Li, John Bradshaw, and Yash Sharma, 2019

We compared the robustness of generative classifiers to adversarial attacks with discriminative classifiers, and verified the hypothesis on a real-world setting.

Toxic Gas Characterization

I worked on a project to characterize toxic gases using a deep learning model and adversarial unsupervised domain adaptation to improve the model's generalization.

Convergence of SGD for Training Neural Networks with Sliced Wasserstein Losses

Eloi Tanguy, 2024

We empirically verified the convergence of SGD for training generative neural networks on 2D distributions and Fashion-MNIST using the Sliced Wasserstein loss.

An End-to-End Transformer Model for 3D Object Detection

Ishan Misra, Rohit Girdhar, and Armand Joulin, 2021

We verified 3DETR's performance on the SUN RGB-D dataset. We trained a RGB variant of 3DETR and compare it to the original model.

Some of my Recent Projects

Audio Visual Transcription

AVT is a tool that lets you subtitle audio/video content in minutes, powered by OpenAI Whisper.

Video Background Removal

A video background removal tool using Mobile SAM (Segment-Anything Model) to automatically segment and remove the background from videos.

Video Object Detection

Implementation of Google Owl-ViT's model for zero-shot object detection in videos using natural language prompts.

Contribution to joliGEN

I integrated edge detection methods into joliGEN (from ControlNET). I also added Segment-Anything Model from Meta for precise masking.

MathViz

MathViz is a Streamlit app that allows you to visualize concepts from Mathematics, Statistics, Machine Learning or Algorithms.

Wikipedia Graph

Scraping Wikipedia to create a structured graph of the concepts around a base word (in French). The output data can be loaded in Gephi to visualize the graph.

Word Ladder Generator

Word Ladder Generator is a game that generates the shortest path between two words by changing one letter at a time. In French.