Hoang Trieu Vy LE

Hoang Trieu Vy LE

2025 –

I am a Research Scientist at Beink Dream in Paris, where I build the harness for end-to-end agentic AI systems. I also work on flow-based image generation and multimodal systems that keep a structured representation of 2D, 3D and CAD scenes.

2024 – 2025

I was a Research Engineer at CEA – DIGITEO Labs in Saclay, where I worked on lightweight learning models for 3D CT reconstruction — building physics-consistent scene representations from sparse, noisy measurements and limited high-dimensional observations.

2020 – 2023

My PhD at ENS Lyon was focused on variations of the Mumford-Shah functional for interface detection in degraded images, bridging proximal algorithms and unrolled deep architectures [take a look] under the supervision of Prof. Nelly Pustelnik and Dr. Marion Foare. Along the way I spent a year at UC Louvain (2020–2021) designing scalable accelerated algorithms for joint image restoration and edge detection on large-scale physics experiment images.

I also taught Introduction to Machine Learning and Optimisation at EPITA, and Machine Learning & Optimization hands-on courses at ENS Lyon.

2015 – 2020

I studied at INSA Toulouse where I earned my Master of Engineering in Applied Mathematics, majoring in Mathematical Modelling and Numerical Analysis.

Research Interests

Agentic systems
- Harnesses for agents: orchestration, tool use, memory, evaluation
- Multimodal LLMs
- Multi-agent systems
- Robotics and embodied agents
Machine learning for perception
- Image processing and computer vision
- Video understanding
- 3D/4D scene understanding
- Generative models and flow matching
- World models, real2sim2real transfer, physical AI
- Reinforcement learning for robot control and human modeling
Inverse problems and optimization
- Non-smooth and non-convex optimization, proximal algorithms
- Unrolled and physics-consistent deep architectures
- Mumford–Shah models, joint restoration and edge detection
Applications
- Computer-aided design and 3D/4D scene understanding
- Tomographic imaging, 3D CT reconstruction

Recent Works

	Zingu — a two-wheeled self-balancing robot A two-wheeled self-balancing robot that picks itself back up after a fall. I'm building it toward an outdoor companion: the robot-dog idea on two wheels, smarter, and open to whatever task I can imagine giving it. Zingu Brain — the perception and decision side — comes next. [Project] [Code]
Gist	Gist Ongoing research on scene understanding through many senses at once. Humans take in vision, touch, sound, a sense of space and a sense of time, and fuse them into thought and reasoning. Sometimes we even dream. Models cannot, and Gist is my attempt to build a multimodal system that perceives, reasons, and maybe one day dreams.
Meepo	Meepo A personal AI assistant. [Code]

Publications

	Unfolded discrete Mumford-Shah functional for joint image denoising and edge detection HoangTrieuVy LE, Marion Foare, Audrey Repetti, Nelly Pustelnik EUSIPCO 2025 [PDF] [Code]
	Embedding Blake-Zisserman Regularization in Unfolded Proximal Neural Networks for Enhanced Edge Detection HoangTrieuVy LE, Marion Foare, Audrey Repetti, Nelly Pustelnik IEEE Signal Processing Letters, 2025 [PDF]
	Proximal Neural Networks based reconstruction for few-view CT applications HoangTrieuVy LE, Caroline Bossuyt, Marius Costin, Hang Wang, Jan De Beenhouwer, Jan Sijbers 14th Conference on Industrial Computed Tomography (iCT 2025), Antwerp, Belgium [PDF]
	Unfolded proximal neural networks for robust image Gaussian denoising HoangTrieuVy LE, Audrey Repetti, Nelly Pustelnik IEEE Transactions on Image Processing, 2024 [PDF] [Code] [BibTeX] @ARTICLE{10630640, author={Le, Hoang Trieu Vy and Repetti, Audrey and Pustelnik, Nelly}, journal={IEEE Transactions on Image Processing}, title={Unfolded Proximal Neural Networks for Robust Image Gaussian Denoising}, year={2024}, volume={33}, number={}, pages={4475-4487}, doi={10.1109/TIP.2024.3437219}}
	Deep image prior for sparse-view reconstruction in static, rectangular multi-source X-ray CT systems for cargo scanning Caroline Bossuyt, Jan De Beenhouwer, Jan Sijbers, Domenico Iuso, Marius Costin, Julie Escoda, HoangTrieuVy LE, Arjan J den Dekker Developments in X-Ray Tomography XV, SPIE, 2024 [DOI]
	Proximal based strategies for solving Discrete Mumford-Shah with Ambrosio-Tortorelli penalization on edges HoangTrieuVy LE, Marion Foare, Nelly Pustelnik IEEE Signal Processing Letters, 2022 [PDF] [Code] [BibTeX] @ARTICLE{9723590, author={Le, Hoang Trieu Vy and Foare, Marion and Pustelnik, Nelly}, journal={IEEE Signal Processing Letters}, title={Proximal Based Strategies for Solving Discrete Mumford-Shah With Ambrosio-Tortorelli Penalization on Edges}, year={2022}, volume={29}, number={}, pages={952-956}, doi={10.1109/LSP.2022.3155307}}
	The faster proximal algorithm, the better unfolded deep learning architecture ? The study case of image denoising HoangTrieuVy LE, Nelly Pustelnik, Marion Foare EUSIPCO 2022 [PDF] [Code] [BibTeX] @INPROCEEDINGS{9909592, author={Le, Hoang Trieu Vy and Pustelnik, Nelly and Foare, Marion}, booktitle={2022 30th European Signal Processing Conference (EUSIPCO)}, title={The faster proximal algorithm, the better unfolded deep learning architecture ? The study case of image denoising}, year={2022}, pages={947-951}, doi={10.23919/EUSIPCO55093.2022.9909592}}
	Fast Proximal Unrolled Algorithms for the Analysis of Piecewise Homogeneous Fractal Images HoangTrieuVy LE, Nelly Pustelnik, Barbara Pascal, Nelly Pustelnik, Marion Foare, Patrice Abry GRETSI 2022 [PDF] [BibTeX] @inproceedings{HoangTVL2022AlgorithmesPR, title={Fast Proximal Unrolled Algorithms for the Analysis of Piecewise Homogeneous Fractal Images}, author={ Hoang T.V Le, Barbara Pascal, Nelly Pustelnik, Marion Foare and Patrice Abry}, year={2022}, url={https://api.semanticscholar.org/CorpusID:253115164} }

Ph.D Thesis

Variations on the Mumford-Shah functional for interface detection in degraded images: from proximal algorithms to unrolled architectures
[PDF]

Supervisors: Prof. Nelly Pustelnik, Dr. Marion Foare

Reviewers: Prof. Caroline Chaux, Assoc. Prof. Matthieu Kowalski

Examiners: Prof. Jean-François Aujol, Prof. Elie Bretin, Prof. Luca Calatroni, Prof. Marc Sebban

Talks

EUSIPCO, Belgrade, Serbia, Aug 2022
Conference on Digital Geometry and Discrete Variational Calculus, Mar 2021

Teaching

Introduction to Machine Learning and Optimisation, EPITA
Machine Learning and Optimization hands-on courses at ENS Lyon

For Fun

Polyverse — a hobby game I build on weekends. A low-poly 3D real-time strategy game in Godot, where time has shattered into crystal shards and each shard is one frozen age of civilization.