Senior AI/ML Engineer & Researcher
Diffusion Models · Robotics · World Models · Physical AI
Enabling World Models and Physical AI development via NVIDIA Cosmos. Making robots see and act in the real world. Innovation & Engineering team @ idealworks GmbH.
I am a Senior AI/ML Engineer at idealworks GmbH in Munich, building the bridge between cutting-edge ML research and production robot systems.
My current work centers on Visual SLAM systems, deploying vision models on edge, managing data infrastructure, leading research projects through Master theses and internships, and using Diffusion Transformers for embodied AI through NVIDIA's Cosmos platform.
I work a transport robot called the iw.hub which is an AMR deployed in over 10 locations with a total of 1200 robots
I hold an M2 in Mathematics, Machine Vision and Machine Learning from École Normale Supérieure Paris-Saclay, one of France's most selective Grandes Écoles.
idealworks GmbH
inmind.ai
BMW Group / idealworks GmbH
École Normale Supérieure Paris-Saclay
Université Paris-Saclay
Holy Spirit University of Kaslik (USEK)
Direct collaboration with NVIDIA to fine-tune and integrate NVIDIA Cosmos data generation and augmentation pipeline into the AMR domain and modalities. Developing Synthetic Data Generation pipeline at scale for production use-cases. Tools used include Grafana, Prometheus, LakeFS, Determined AI for GPU inference job orchestration, FastAPI for API development, and Docker for containerization.
NVIDIA Cosmos Webpage→Collaborated with the NVIDIA cuVSLAM team to publish a whitepaper on cuVSLAM benchmark and integration in industrial settings where we showcased the performance of cuVSLAM in comparison to other SLAM algorithms. Frameworks used include ROS2, CUDA and NVIDIA Omniverse for Simulation testing.
cuVSLAM→ arXiv→Designed and deployed a hybrid LiDAR + Visual SLAM system for our iw.hub fleet after conducting an elaborate research and investigation into idealwork's LiDAR based localization system shortcomings. The system is currently in pilot mode and is expected to be deployed and scaled to full on large-scale production system by end-of-year. Designed a benchmarking framework deployed on-site and on AWS for system comparison and agile updates. Frameworks used include ROS2, CUDA, FastAPI, AWS EC2, AWS Lambda, AWS Batch, AWS ECR,Github Actions, Bazel.
Built and deployed a modular people anonymization pipeline for autonomous mobile robots in production. The pipeline support modular model support with automatic benchmarking and selection. Deployed on-site and on AWS with Lambda, EC2, Batch and S3. CI/CD with Github Actions. On-site deployment uses Docker and Celery for distributed processing.
Built and deployed a robot motion prediction pipeline on-edge on NVIDIA Jetson Xavier device for on-edge collision avoidance in production. Technologies for the core were a YOLO model for object detection and an Unscented Kalman Filter for motion estimation and tracking. Deployment was done using ROS, Docker, ONNX and TensorRT.
Built infrastructure and pipelines for robot data collection and processing in production which includes people anonymization, ROS2 bags handling, version control and data storage. Building large industrial datasets for training world models and other perception models. Built an extension on top of Rerun for custom data visualization, analysis and annotation. Initiated the Data Driven Decision Making (DDDM) framework for better problem understanding and solutions delivery. Frameworks used include ROS2, DVC and LakeFS for data version control, NAS and AWS S3 for data storage.
Collaboration with NVIDIA · First Author
A comprehensive benchmark of visual odometry and Visual SLAM systems for mobile robot navigation in real-world logistics environments. Multiple VO approaches are evaluated across controlled translational, rotational, and mixed-motion trajectories, as well as a large-scale ~1.7 km production facility dataset, using Absolute Pose Error against Vicon motion-capture and LiDAR-SLAM ground truth. Results show that a hybrid stack combining the cuVSLAM front-end with a custom SLAM back-end achieves the strongest mapping accuracy, motivating its deeper integration as the core VO component. The stack is further validated through deployment on an NVIDIA Jetson platform.
Extended Abstract · First Author
Addresses key limitations of imitation-learning-based path planning for AMRs—limited generalization to unseen environments and fragile demonstration collection—through two main contributions: an overhauled ROS 2 annotation tool and a novel training strategy that integrates diffusion-based augmentation into baseline behavioral cloning models. Evaluated via ablation studies on a provided expert-demonstration dataset, the enhanced approach outperforms state-of-the-art methods, achieving 39.1% lower Absolute Pose Error (APE) and 33.5% lower Fréchet Inception Distance (FID) while requiring 93.8% fewer trainable parameters, and attains diffusion-level generalization while preserving real-time, on-edge inference properties.
IEEE IPAS · Second Author
Introduces a more accessible and flexible approach to AMR path planning through sketch-guided imitation learning, where non-technical users simply draw the desired navigational path on a provided 2D map to teach U-net models path planning behaviors—replacing traditional reward-based learning or costly hardware-dependent techniques. A novel evaluation framework combining metrics from image generation and robotics is proposed, and the approach is integrated into an end-to-end robotics stack to demonstrate practical usability.
Open-source work and research implementations. View all repositories on GitHub.
End-to-end Sketch-Guided Path Planning through Imitation Learning for Autonomous Mobile Robots
Implementation of the publication "End-to-end Sketch-Guided Path Planning through Imitation Learning for Autonomous Mobile Robots".
Semantic segmentation for autonomous robot navigation in urban environments
Pixel-level semantic segmentation enabling robots to identify safe sidewalk areas from RGB camera feeds. Complete pipeline from data labeling through training to real-time inference with a custom 276-image dataset.
Multi-Agent Reinforcement Learning benchmarking for decentralized autonomous driving
Benchmarking suite for evaluating decentralized multi-agent RL algorithms in high-density autonomous driving scenarios using highway-env simulation.
GANs for data augmentation and image-to-image translation
Tackling data scarcity through GAN-based synthetic image generation and CycleGAN for unpaired image-to-image translation across visual domains.
Regression analysis for predicting professional athlete market values
End-to-end ML project predicting football player transfer values. Demonstrates reproducible ML workflows with DVC for experiment tracking and Dockerized execution.
Healthcare analytics: classification model for stroke risk prediction
Classification project for predicting stroke risk using patient health data. Covers EDA, class imbalance handling, and multi-classifier comparison.
Utility scripts for verifying GPU acceleration across TensorFlow 2 and PyTorch environments.
View on GitHubCustomized Neovim configuration built on AstroNvim, optimized for Python, C++, and CUDA development.
View on GitHubData and model version control using DVC in MLOps — setup guides, remote storage, and Python API integration.
View on GitHubProfessional LaTeX templates for academic publications and conference papers (IEEE format presets).
View on GitHub