Senior AI engineer

3D vision, from geometry to generative models.

Senior AI engineer with 5+ years in 3D computer vision, neural rendering and generative AI. I work end to end. I build the training data, design and train the model, then ship it as an inference service the product calls.

Get in touch

01 / Experience

Experience.

Production work across 3D reconstruction, generative AI, model training and deployment.

Sept 2022 — PresentFull-time

Deep Learning Specialist — 3D Vision & Generative AI

Dassault Systèmes · Aix-en-Provence, France

Current

Own the stack from data to deployment, across both classical geometry and large generative models.

Generative & multimodal AI

Built a VLM on top of a fine-tuned Qwen LLM that takes 3D tokens as an extra modality. It answers questions about a shape and produces embeddings that drive text-based 3D editing.
Trained a large latent diffusion transformer (DiT) for image-to-mesh generation. Started on Objaverse, then scaled to TexVerse, ABO and internal CAD data.
Trained a point cloud VAE that encodes 3D shapes into a compact latent space. This is the latent space the DiT runs on.
Added physical constraints to the diffusion process using physics solver gradients on the SDF, so generated objects hold up physically as well as visually.
Fine-tuned DINOv2 image encoders on industrial CAD imagery.

3D reconstruction & rendering

Led digital twin reconstruction projects for industrial sites using photogrammetry and neural rendering.
Built SfM/MVS pipelines with COLMAP and Ceres, handling 1,000+ image captures with custom calibration and distributed densification.
Replaced classical feature extraction and matching with SuperPoint and LightGlue. Reconstruction dropped from hours or days to minutes with no loss of accuracy.
Worked closely with the team on NeRF and Gaussian Splatting to extend the photogrammetry pipeline for better visual fidelity.
Trained segmentation and detection models (Mask R-CNN) for object isolation and background removal, running in real time on GPU.

Training data

Built synthetic data pipelines in Blender and UE5 with domain randomisation, which cut our dependency on outside datasets.
Wrote an auto-annotation pipeline that renders multiple views of a 3D object, computes DINOv2 and VLM features, then writes captions and tags. This is what taught the LLM the new 3D modality.
Handled data compliance and licensing across all datasets.

Training, deployment & collaboration

Ran training on large multi-GPU NVIDIA clusters with mixed precision and sharded data parallel.
Shipped models as Docker GPU inference services with model versioning, monitoring and latency tuning.
Worked closely with the UI/UX and front-end teams on the inference REST API. We agreed the request and response format, async job handling, progress updates for long generations, and fallback behaviour, then iterated through UX testing.
Mentored two interns (2023 and 2024) and worked across brands, R&D and deployment teams.

Mar 2022 — Sept 2022Internship

Computer Vision Engineer

Dassault Systèmes · Aix-en-Provence, France

Photogrammetry and deep-learning-assisted 3D reconstruction R&D.

R&D in photogrammetry and deep-learning-assisted 3D reconstruction.
Prototyped preprocessing and calibration methods that went into production.

Feb 2021 — Jan 2022Study Project

Developer

DXOMARK · Boulogne-Billancourt, France

Real-time camera quality visualisation tooling.

Built Qt interfaces with OpenGL shaders for real-time camera quality visualisation.

Nov 2020 — Jan 2021Internship

Back-End Developer

SNCF Réseau · Saint-Denis, France

REST API design for railway infrastructure management.

Designed REST APIs with OpenAPI/Swagger documentation for railway infrastructure management.

02 / Selected work

Selected work.

Systems I have built across reconstruction, multimodal learning, 3D generation and rendering.

Overview

Developed and deployed a comprehensive pipeline for reconstructing large-scale outdoor environments including terrains, factories, and industrial sites. The system enables the creation of high-fidelity digital twins for simulation and industrial use cases.

Challenge

Large outdoor scenes present unique challenges including varying lighting conditions, scale consistency across vast areas, and computational efficiency for processing thousands of images.

Solution

Implemented an optimized SfM/MVS pipeline using COLMAP with custom camera calibration routines and distributed point cloud densification. Developed intelligent image selection algorithms to reduce redundancy while maintaining coverage.

Impact

Delivered digital twins of industrial facilities for virtual inspection, planning and simulation. Cut on-site survey time while keeping production accuracy.

Key Highlights

Optimized camera calibration and point cloud densification
High-accuracy results at scale for digital twin applications
Handles 1,000+ image captures per scene

Overview

Advanced reconstruction system specifically designed for indoor environments and individual objects. Combines traditional photogrammetry with modern neural rendering techniques to handle challenging indoor scenarios.

Challenge

Indoor scenes feature complex lighting, reflective surfaces, occlusions, and repetitive textures that confuse traditional reconstruction methods. Scale recovery is particularly challenging without natural features.

Solution

Integrated Gaussian Splatting and neural rendering techniques with traditional photogrammetry. Developed custom preprocessing for handling reflections and implemented scale recovery using learned priors and geometric constraints.

Impact

Achieved photorealistic reconstructions of indoor spaces with accurate geometry. System now used for interior design visualization and virtual staging applications.

Key Highlights

Enhanced visual fidelity and rendering performance
Accurate geometry with scale consistency
Real-time rendering capabilities for interactive visualization

Overview

An end-to-end generative 3D stack I own myself. A point cloud VAE compresses geometry into a compact latent space, and a large latent diffusion transformer trained on that space generates meshes conditioned on an image, a text prompt or existing 3D geometry.

Challenge

Generating 3D from 2D is ill-posed. The latent space has to stay compact enough to train a diffusion model on, while still holding enough detail that the decoded mesh is usable. And no dataset existed that paired 3D geometry with the kind of description the model needed.

Solution

Trained the VAE first, then the diffusion transformer on its latent space. Started on Objaverse and scaled to TexVerse, ABO and internal CAD data. Built the supervision myself with an automatic multi-view captioning pipeline when the annotations turned out to be too thin to train on.

Impact

Cut 3D content creation from hours to minutes, and let non-experts create geometry from a description instead of modelling it by hand.

Key Highlights

Point cloud VAE plus latent diffusion transformer, trained from scratch
Conditioned on image, text or existing 3D geometry
Trained on Objaverse, scaled to TexVerse, ABO and internal CAD

Overview

A vision-language model built on a fine-tuned Qwen LLM that accepts 3D geometry tokens alongside image and text embeddings. It answers questions about a shape and produces embeddings that condition a diffusion model to edit that shape from a prompt.

Challenge

A language model has never seen geometry. Getting it to reason about 3D means deciding how a shape becomes tokens it can attend over, and adding that modality without destroying the model’s language ability.

Solution

Designed the tokenisation and the adapter layers that project geometry into the language model’s embedding space, then picked a fine-tuning strategy that added the modality while preserving what the model already knew. Built the training corpus with an automatic annotation pipeline: render multiple views, compute DINOv2 and VLM features, write captions and tags.

Impact

Shapes became editable by description rather than by manipulation, and the model can answer questions about geometry it is given.

Key Highlights

3D tokens injected into a fine-tuned Qwen LLM
Embeddings condition downstream diffusion for prompt-driven editing
Training supervision generated automatically, since none existed

Overview

Generative models produce shapes that look right and fall over. This work adds physical constraints to the diffusion process itself, using gradients from a physics solver operating on the signed distance field.

Challenge

A purely visual objective has no notion of whether an object can stand, hold together, or bear load. Adding that signal without breaking the image, text and geometry conditioning is the hard part.

Solution

Used solver signals on the implicit surface as guidance during sampling, feeding physical information back through the generative process alongside the existing conditioning.

Impact

Generated geometry satisfies physical constraints as well as the prompt, which matters as soon as the output is meant for simulation or manufacture rather than a render.

Key Highlights

Physics solver gradients on the SDF used as guidance
Physical validity alongside image, text and geometry conditioning
Worked across solver output, implicit surfaces and finite elements

Overview

Developed a specialized AI segmentation system to cleanly isolate objects of interest from complex backgrounds, significantly improving the quality of downstream 3D reconstruction tasks.

Challenge

Cluttered backgrounds introduce noise and artifacts in reconstruction. Traditional methods struggle with complex scenes, transparent objects, and fine details like hair or thin structures.

Solution

Trained a custom segmentation AI model using a combination of real and synthetic data. Implemented multi-scale feature extraction and boundary refinement modules to capture fine details. Model optimized for real-time inference.

Impact

Improved reconstruction quality by eliminating background noise. Reduced post-processing time. System now integral part of production reconstruction pipeline.

Key Highlights

Improved reconstruction quality and robustness
Removed background noise prior to photogrammetry stages
Real-time inference on GPU

Overview

Implemented a high-performance ray tracing engine from scratch using CUDA and modern rendering techniques. The system supports physically based materials, global illumination, and real-time visualization.

Challenge

Ray tracing is computationally intensive. Achieving real-time performance requires careful optimization of ray-primitive intersection tests, memory access patterns, and parallel workload distribution.

Solution

Implemented BVH acceleration structures with SAH-based construction. Optimized CUDA kernels for coherent ray batching and memory coalescing. Developed hybrid rasterization-ray tracing pipeline for interactive frame rates.

Impact

Interactive frame rates for complex scenes with global illumination, used for high-quality visualisation of digital twins and design validation.

Key Highlights

Real-time visualization and rendering components
Physically based rendering principles
Optimized for both quality and performance

Overview

Created comprehensive synthetic data generation pipelines that produce photorealistic training data for various computer vision AI tasks. System employs domain randomization and physically-based rendering to maximize sim-to-real transfer.

Challenge

Acquiring large-scale labeled real-world data is expensive and time-consuming. Synthetic data must be diverse and realistic enough to train AI models that generalize to real scenarios.

Solution

Built automated pipelines in Blender and Unreal Engine 5 with extensive domain randomization. Implemented PBR materials, realistic lighting, and physics simulation. Developed tools for automatic annotation and quality control.

Impact

Removed the dependency on outside datasets, cutting data acquisition cost and licensing risk. AI models trained on synthetic data achieve performance comparable to real-data trained models.

Key Highlights

Employed domain randomization and PBR materials
Closed the sim-to-real gap for AI training
Fully automated pipeline with quality control

03 / Reading

Reading.

The research behind my work and the ideas I continue to track.

The architecture family behind the point cloud VAE and latent diffusion transformer I train, and the work I follow most closely.

04 / Background

Background.

Education, technical toolkit and languages.

Education

2019 — 2022

MSc Computer Engineering

EPITA — École d’Ingénieurs en Informatique

Computer Vision, Machine Learning, Image Synthesis

2021 — 2022

Master 2 — Ingénierie Informatique (IMA)

Sorbonne Université

Image Processing, Computer Vision, Machine Learning, Image Synthesis

Languages

Technical toolkit

3D vision, from geometry to generative models.

Deep Learning Specialist — 3D Vision & Generative AI

Generative & multimodal AI

3D reconstruction & rendering

Training data

Training, deployment & collaboration

Computer Vision Engineer

Developer

Back-End Developer

Large-Scale Outdoor Scene Reconstruction

Indoor Scene & Object Reconstruction

Image and Text to 3D Mesh

Teaching an LLM a 3D Modality

Physics-Guided 3D Generation

Robust Background Removal for Object Reconstruction

GPU Ray Tracing & Rendering

Synthetic Data Generation for AI Training

Hunyuan3D-Buffalo 1.0: A Unified Multimodal Model for Scalable 3D Generation, Understanding, and Editing

Axolotl3D: a Unified Framework for Faithful 3D Shape Completion

LATTICE: Democratize High-Fidelity 3D Generation at Scale

NANO3D: A Training-Free Approach for Efficient 3D Editing Without Masks

Cube: A Roblox View of 3D Intelligence

TRELLIS: Structured 3D Latents for Scalable 3D Generation

Hunyuan3D 2.0: High-Resolution Textured 3D Asset Generation

Scalable Diffusion Models with Transformers (DiT)

Denoising Diffusion Probabilistic Models

High-Resolution Image Synthesis with Latent Diffusion Models

Flow Matching for Generative Modeling

MSc Computer Engineering

Master 2 — Ingénierie Informatique (IMA)

Programming Languages

Deep Learning & AI

3D Vision & Graphics

Simulation & Synthetic Data

DevOps & MLOps

Cloud & Storage