Anar Nurizada

About

Robotics/ML engineer who benchmarks and adapts open-source robot action models, building synthetic datasets in Isaac Sim and modifying vendor code to scale training across multiple GPUs. Also built a production-style computer vision pipeline at Zortag—training a tiny YOLO-based classifier for fake-vs-real label detection and deploying it in a real-time iOS app with additional display/spoof detection.

Experience

Education

Key Strengths

• Benchmarked and fine-tuned open-source robotics action models (GR00T vs pi0) to 90%+ success on unseen pick-and-place dataset
• Built synthetic robotics dataset pipeline in Isaac Sim despite poor documentation and instability (auto-restart via bash script)
• Rewrote NVIDIA GR00T training code to enable multi-GPU training
• Built end-to-end CV product pipeline: dataset creation, YOLO-based classifier training, and iOS app for real-time inference
• Achieved near-100% label authenticity classification using smallest YOLO model (~5MB) enabling fast on-device inference
• Used systematic augmentation testing and ablation studies to improve model performance efficiently
• Integrated ROS2 with MoveIt2 for motion planning around randomized obstacles for dataset generation
• Fine-tuned and benchmarked robot action models; achieved 100% vs 93% performance (pi0 vs GR00T) on same pick-and-place dataset
• Deep debugging and adaptation of research code to run on multi-GPU setups
• Iterative robotics dataset design (camera count/placement, episode length) with retraining and inference loops
• Automated dataset creation using a robot arm + camera to avoid manual labeling/video capture
• Improved YOLO fake-label detector to 100% accuracy across multiple mobile devices via augmentation ablation study
• Strong understanding of embodied AI constraints: compute scaling, dataset collection time, and out-of-distribution generalization limits
• High-impact decision-making under time pressure (selected feasible approach in 48-hour hackathon; team won 1st place)