New York, New York, United States
♦ Google scholar: https://scholar.google.com/citations?user=y_uRucIAAAAJ ♦ Github: https://github.com/gsicard
♦ Part of a pluridisciplinary research team comprised of machine learning researchers, neuroscientists and hardware engineers within Meta Reality Labs Neural Interfaces group (ex-CTRL-Labs) ♦ Developing ML models (timeseries, language models...) to decode handwriting from wrist-based EMG (electrical signals from forearm muscles), investigating model error modes and developing cross-stack scientific solutions (data, modeling...) ♦ Reduced our top 2 model error categories rate by 56% and 70% with targeted modeling solutions for the Neural Handwriting Early Access Program release ♦ Characterized our top error category, influencing leadership roadmap, implemented the corresponding offline metrics and led a cross-stack mitigation task force ♦ Improved our main model’s GPU efficiency by 70% with efficient vectorization, data bottleneck reductions and optimized sampling ♦ Promoted usage of agentic systems for research in the team (skills, practices...) Neural Handwriting was recently released as an early access feature for the Meta Ray Ban Displays with the Meta Neural Band and presented during a live demo by Mark Zuckerberg at 2025 Connect. The foundations of the EMG work were published in Nature.
Tech Lead, Individual Contributor ♦ Advised machine learning researchers and domain experts on ML model research and domain-specific evaluation and product managers on the ML landscape ♦ Reduced our production ML model error rate by 30% on targeted test sets by researching novel model architectures inspired by recent scientific publications ♦ Provided insights into our model predictions to our customers after researching data and ML model explainability methods specific to our architecture
♦ Managed a team of 4 research scientists and engineers focused on ML innovation ♦ Reached over 90% recall on targeted ML adversarial use cases by investigating domain-specific data characteristics to improve our ML model architecture ♦ Provided a product feature aimed at challenging sample groups, reaching 95% recall by researching explainability methods to target data-specific model weaknesses
♦ Reduced our overall production ML model error rate by 15% by improving our machine learning architecture (Python, TensorFlow, PyTorch) ♦ Improved ML model training speed by 3x by developing a custom data format with Python and PyTorch interface, published on GitHub, to support 100+ million training samples (https://github.com/gsicard/syrah) ♦ Co-authored research articles and presented at an academic conference
♦ Comparison of two state-of-the-art function approximation algorithms for multiple motor control models learning in humanoid robotics : LWPR (Locally Weighted Projection Regression) and XCSF (eXtended Classifier System for Function approximation). ♦ Software development and experimentations on the iCub humanoid robot.
♦ Using the XCSF algorithm (eXtended Classifier System for Function approximation) and stereo-vision to learn motor control models for humanoid robots
♦ Object recognition and tracking for robotic navigation