Sourabh Singh

Designed a deep learning video pipeline for AAOS on Snapdragon devices (Android/8155p). Engineered seamless integration of UVC (USB Video Class) with OpenCV on AAOS, reducing implementation time by 30%. Utilized YOLOv5 with precision, achieving 92.5% accuracy for real-time object detection on Android at a seamless 20 FPS, augmented by leveraging SNPE (Snapdragon Neural Processing Engine). Led the development, maintenance, and strategic leadership of a Deep Learning SDK, contributing to a 30% improvement in the SDK. Contributed to the development of two applications, achieving State-of-the-Art (SOTA) performance in both accuracy and inference time. Documented and oversaw maintenance for two devices, namely RZV2L and RZV2M.

Developed and maintained a custom-trained TAO model, seamlessly integrating it with RabbitMQ as a message broker, resulting in a 40% reduction in data extraction time from the server. Integrated the post-processing of data into a Peewee database, concurrently establishing endpoints to facilitate customer access to CSV files. Innovated by creating custom metrics, such as warehouse-util and warehouse-eff, to provide nuanced insights and better align with organizational goals. Managed end-to-end data processes, including cleaning, analysis, and construction of datasets from 32+ IoT devices, followed by the training of ML models and pipeline. Implemented Flask APIs to seamlessly interact with the trained model, ensuring efficient integration into applications. Attained 96% accuracy with a 0.2-second inference time, demonstrating the model’s efficiency and precision. Engineered and maintained a cutting-edge system for traffic monitoring in EVMs, utilizing AWS Rekognition to enhance accuracy by 25%. Devised an audio analysis solution using Librosa and YAMNet to accurately count beeps, resulting in a 20% improvement in precision. Effectively streamlined and optimized data annotation tasks through AWS Sagemaker, improving workflow efficiency and accelerating annotation speed by 30%. Managed the deployment of FasterRCNN, designed an optimized database, and developed a responsive Flask API, achieving a 20% increase in model performance and a 30% reduction in API response time. Guided the training of a highly accurate FasterRCNN model and an efficient OpenCV algorithm for defect detection, resulting in a 15% improvement in defect identification accuracy. Successfully deployed the solution on AWS using Sagemaker and Lambda.

Implemented OCR utilizing Google Vision API and Tesseract, achieving a 95% accuracy in text extraction. Led initiatives to optimize costs, resulting in a 20% reduction through efficient Computer Vision techniques, spell checking, and data cleaning before inference. Fine-tuned thresholds, improving overall efficiency and reducing costs by 15% in calls to the Google Vision API.