Alexander Soare
Product Engineer
Summary
🤩
RockstarAlexander Soare is a Senior AI Research Scientist focused on embodied AI and robotics, bringing a decade of experience bridging research, product, and hands-on systems engineering. He currently leads agile, hardware-integrated research at Cobot after helping found Hugging Face’s LeRobot project, where he contributed to diffusion policies and multi-image observation encoders. An ex-physicist who built perception stacks for logistics robots, he has deep expertise in 3D pose, multi-view RGBD, and adapting transformer architectures (notably work on ViT/TNT positional embedding resizing and a Nested Transformer implementation in timm). He pairs rigorous academic training in quantum physics with pragmatic product and ops experience from startups to enterprise, enabling rapid prototyping through production. Known for sharing work openly—running hackathons and podcast episodes—he blends curiosity-driven research with measurable impact in real-world robotic systems.
10 years of coding experience
Github Skills (11)
diffusion-models10
pytorch10
machine-learning10
image-processing9
python9
data-normalization9
transformers9
computer-vision9
robotics8
neural-network5
convolutional-neural-networks5
Programming languages (7)
Github contributions (5)
huggingface/lerobot
May 2024 - Oct 2024
🤗 LeRobot: Making AI for Robotics more accessible with end-to-end learning
Role in this project:
ML Engineer
ML EngineerContributions:330 reviews, 120 PRs, 64 pushes in 5 months
Contributions summary:Alexander contributed to the development of the LeRobot project, specifically within the context of diffusion policies. The user's commits focused on refining the multi-image observation encoder, incorporating normalization techniques, and ensuring compatibility with different image shapes. Additionally, the user made substantial changes to the training script, integrating new features and addressing issues with online training and evaluation. These modifications indicate an effort to improve the model's performance and streamline the training process.
airobotics
huggingface/pytorch-image-models
May 2021 - Nov 2021
The largest collection of PyTorch image encoders / backbones. Including train, eval, inference, export scripts, and pretrained weights -- ResNet, ResNeXT, EfficientNet, NFNet, Vision Transformer (ViT), MobileNetV4, MobileNet-V3 & V2, RegNet, DPN, CSPNet, Swin Transformer, MaxViT, CoAtNet, ConvNeXt, and more
Role in this project:ML Engineer
ML EngineerContributions:24 commits, 7 PRs, 29 comments in 5 months
Contributions summary:Alexander's contributions focused on enhancing positional embedding resizing within the `pytorch-image-models` repository, specifically for Vision Transformer and TNT models. They extended the resizing functionality to accommodate non-square grids and incorporated it into the TNT model. Furthermore, the user is actively working on a Nested Transformer (NesT) implementation within the repository, including the conversion of weights from the original Jax implementation. This work suggests a focus on model architecture modifications and adaptation of pre-trained weights.
efficientnetinferencemobilenetpytorchresnet
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.