Binesh Thoppil

Led a team of 3 building the backend for the visionbot platform, handling custom object recognition, segmentation, and predictive models on over 400 camera feeds simultaneously at 10 FPS. Developed a CNN-based custom object detection model with 75% mAP using Tensorflow. Increased inference rate by 5x using CUDA library (GPU Acceleration). Deployed modules in Docker containers that auto-spawned on AWS instances, reducing downtime by 95%. Developed ML models and data collection pipelines for use cases spanning construction-sites, logistics, and inventory management. Built a model for human pose estimation (83% mAP) and implemented object tracking across frames and cameras. Optimized model weights using Nvidia TensorRT, increasing inference rate by 200% and reducing memory usage by 40%. Wrote multi-threaded/multi-process Python programs and used APIs to run modules as micro-services. Implemented load sharing and auto-scaling algorithms for new EC2 instances. Worked with scalable cloud services including Amazon S3 and Rackspace. Built edge-based on-premise data collection points using Raspberry Pi, ESP32, and Nvidia Jetson Nano. Collaborated on feature engineering and labeling, reducing turnaround time by >90%. Took complete ownership of developed modules. Maintained various documentation (design, user, maintenance, project, release). Presented solutions and functionality to stakeholders and clients.