Kaviya Ns

Built and maintained the complete machine learning training data pipeline for exercise-equipment detection by annotating datasets in Roboflow with select and segment tools, testing multiple augmentations, and evaluating YOLO object detection models to improve weight detection and classification robustness. Designed an automated computer vision video-annotation pipeline using Segment-and-Track Anything (SAM) to generate per-frame masks for dumbbells and kettlebells, automatically sort outputs by exercise type and weight class, and reduce manual annotation effort. Extracted SMPL-based point cloud meshes from posture videos across 12 exercises; optimized point density (pruning non-joint regions) to balance BiLSTM inference latency against classification accuracy.

AI Engineer Intern at Smartan FitTech (2025-07 – 2025-12)

• Built and maintained the complete machine learning training data pipeline for exercise-equipment detection by annotating datasets in Roboflow with select and segment tools, testing multiple augmentations, and evaluating YOLO object detection models to improve weight detection and classification robustness.
• Designed an automated computer vision video-annotation pipeline using Segment-and-Track Anything (SAM) to generate per-frame masks for dumbbells and kettlebells, automatically sort outputs by exercise type and weight class, and reduce manual annotation effort.
• Extracted SMPL-based point cloud meshes from posture videos across 12 exercises; optimized point density (pruning non-joint regions) to balance BiLSTM inference latency against classification accuracy.
• Integrated depth data from a Luxonis OAK-D depth camera with SMPL point clouds to resolve 3D joint movement (e.g., forward elbow drift) not observable in 2D, and benchmarked inference performance with vs. without depth input.

M.Sc. in Data Science – PSG College of Technology (2022-08 – 2027-05)