Researched and evaluated hand-object contact detection models to maximize accuracy under occlusion and viewpoint variation. Collected, annotated, and curated custom healthcare-related video datasets and built data preparation/quality assurance workflows for supervised learning. Developed and benchmarked video action recognition systems using pretrained backbones (SlowFast), transformer-based models (VideoMAE), and self-supervised learning (V-JEPA), including hyperparameter optimization and failure-case analysis to improve robustness under domain constraints. Built a V-JEPA-based feature fusion pipeline combining video embeddings with pose-derived kinematic features and temporal metadata for healthcare classification under limited-label settings. Developed pose estimation pipelines to extract and analyze kinematic features from body keypoints such as knee/hip flexion angles.

Video Analytics Intern at Kabam Robotics

September 1, 2024 - March 1, 2025

Developed an end-to-end edge AI video analytics system for security applications, including training YOLO object detection models on custom annotated datasets and integrating them into an existing backend inference pipeline. Prototyped and evaluated edge deployment on NVIDIA Jetson using Frigate and DeepStream to assess real-time on-device analytics feasibility under performance/scalability constraints. Improved evaluation efficiency by building a reproducible containerized pipeline (FFmpeg and MediaMTX) to simulate RTSP streams for hardware-independent inference testing. Built a real-time event-driven alerting architecture using MQTT to transmit edge inference outputs to backend services for downstream workflows including automated S3 storage and email notifications. Experimented with NVIDIA DeepStream multi-stream inference using GStreamer-based pipelines to reduce deployment costs by consolidating inference workloads at the edge and improving GPU utilization.