Neural Video Compression using autoencoders
Imagine describing a complex image using only 10 keywords, after that, someone else try to redraw the image leveraging the 10 keywords.
01
About Me
Computer Vision and Deep Learning Engineer — ESIR, Université de Rennes.
I build intelligent systems that understand, reconstruct, and analyze visual data. My work focuses on image processing, computer vision, information theory, and deep learning — with strong interests in visual reconstruction, inverse problems, and generative models.
What drives me: bridging theory and practice by combining mathematical modeling, research-driven experimentation, and scalable high-performance implementation.
Currently seeking a PhD or full-time position in computer vision, image processing, and machine learning.
Outside of work: basketball, football, and following the latest in AI research.
40%Video compression achieved
33dBPSNR quality metric
85%Medical AI accuracy
<10msReal-time frame rate
02
Research Interests
My work sits at the intersection of mathematics, signal processing, and deep learning. Here are the areas I think about most deeply.
Computer Vision
Visual understanding, recognition, and reconstruction — from classical geometry to modern deep learning pipelines.
Generative Models & Diffusion
Exploring score-based models, diffusion processes, and VAEs for image synthesis and inverse problem solving.
Information Theory
Rate-distortion theory, entropy coding, and the mathematical foundations of image and video compression.
Image & Video Processing
Neural video compression, codec design (JPEG, H.264, HEVC), and real-time processing pipelines.
3D Reconstruction & Multi-view Geometry
Learning-based and geometry-based methods for recovering 3D structure from 2D observations.
Inverse Problems
Denoising, deblurring, super-resolution — using optimization and deep priors to recover signals from degraded observations.
GPU Computing & Optimization
CUDA kernels, parallel algorithms, and high-performance pipelines for compute-intensive visual workloads.
Medical Image Analysis
Segmentation, classification, and computer-aided diagnosis applied to clinical imaging datasets.
03
Projects
Deep Learning · Multimodal
CSIRO — Image2Biomass
Multimodal biomass estimation from aerial imagery combining DINOv2 and tabular metadata.
Medical AI · Classification
Pneumonia CAD
ResNet18 model for computer-aided diagnosis on Chest X-Ray images — 84.93% accuracy.
Medical AI · Segmentation
Cell Segmentation & Tracking
U-Net architecture for pixel-level segmentation on medical imaging datasets.
3D · Multi-view Geometry
Multi-view 3D Reconstruction
Pipeline combining geometry-based and learning-based methods to recover 3D structure from 2D images.
Physics · Neural Networks
Gravitational Network
MLP approximating multi-body gravitational potential in N dimensions.
Inverse Problems · Vision
Image Denoising
Autoencoder architecture for image denoising and signal reconstruction.
Time Series · RNN
Time-Series Forecasting
LSTM model for many-to-one prediction on a simulated sine wave.
NLP · Classification
GRU Text Classifier
Reuters newswire topic classification across 46 categories using GRU.
Computer Vision · CNN
CIFAR-10 Classification
CNN trained from scratch on CIFAR-10 with augmentation and TensorBoard — 70.02% accuracy.
3D Graphics · C++
OpenGL Renderer
Rendering pipeline for 3D models with OpenGL, Blender, and Assimp.
3D Printing · Engineering
3D Reverse Engineering
3D shape reconstruction from GCode with geometric analysis and comparison.
04
Experience
Mar — Sep 2025
Dynamixyz / Take Two Interactive
Neural Video Compression for HMC
- Studied state-of-the-art neural video compression techniques.
- Implemented CNNs and optimized video processing pipelines.
- Achieved 40% compression rate with PSNR of 33 dB.
Oct 2024 — Mar 2025
IRISA & Université de Rennes
3D Reconstruction from G-Code
- Developed a geometric reconstruction pipeline from structured data.
- 3D shape modeling with spatial constraints and optimization.
- Experimental validation on complex and noisy data.
Jun — Oct 2024
InterDigital R&D
Real-Time C++ Video Player
- Integrated DeckLink SDK with professional video capture systems.
- Converted YUV v210 → UYVY for high-performance real-time streaming.
- Achieved 3–10 ms per frame with 100% frames captured.
05
Skills
Python
C++
CUDA
Matlab
PyTorch
TensorFlow
OpenCV
Open3D
ONNX
Git
Linux
Docker
Languages
Deep Learning
PyTorchTensorFlow / Kerasscikit-learn
CNNU-NetRNN / LSTM / GRU
TransformersDINOv2ResNet
AutoencodersDiffusion Models
Vision & Video
OpenCVOpen3DJPEG
H.264HEVCOpenGL
BlenderAlbumentationsONNX
Tools & Infra
GitLinuxDocker
TensorBoardKaggle
Vertex AINumPy / Pandas
06
Certifications
IBM
Deep Learning & Reinforcement Learning
IBM Professional Certificate
↗
GCP
MLOps — Vertex AI
Google Cloud
↗
CG
C++ Programming
CodinGame
↗
DL.AI
AI for Medical Diagnosis
Deeplearning.AI
↗
08
Blog
Thoughts on computer vision, deep learning, and research — written to share what I learn. Also published on Medium.
07
Get in Touch
I'm actively looking for a PhD position or full-time role in computer vision and ML. Let's talk.