Bholay nath singh

Developed an automated car damage detection system using OpenCV, YOLOv8; deployed via FastAPI endpoints; optimized inference speed by 30%, enabling real-time panel-level analysis from morning and evening vehicle images. Implemented semantic segmentation techniques alongside object detection to localize and delineate damaged regions with greater precision, improving detection accuracy by 25% under varying lighting and occlusion conditions. Integrated LLMs and GenAI models for damage severity classification and contextual insights, reducing manual inspection efforts by 40%. Designed and deployed a comprehensive reporting pipeline for service personnel, summarizing damage type, intensity, and recommended actions; end-to-end workflow efficiency improved by 35%. Developed a real-time video analytics system using RTSP streams forwarded to DigitalOcean, leveraging NVIDIA DeepStream on Jetson Orin for person detection, tracking, age and gender classification, and behavioural analysis. Implemented region-based dwell and exit detection, and demographic insights using NVIDIA pre-trained models and tuned Vision-Language Models (VLMs) for efficient inference on AWS Lambda, SageMaker, and EC2 to ensure scalable, low-latency processing. Aggregated detection metrics at hourly and daily intervals to generate actionable KPIs for retail dashboards, enhancing visibility of consumer flow and engagement zones. Implemented real-time alerting via Telegram and email to store managers for rapid response to crowding or long wait times.

Designed and implemented end-to-end ML pipeline for document classification and alert analysis, incorporating data preprocessing, feature engineering, model training, and evaluation to ensure high accuracy and production reliability. Deployed scalable inference architecture on AWS using Lambda functions and EC2 instances, enabling low-latency, event-driven processing and seamless integration with existing business systems. Working on Vera AI, a multilingual document intelligence system leveraging advanced OCR engines and Vision-Language Models (VLMs) for robust text extraction and translation across complex document formats. Improved extraction and translation accuracy by 30% through advanced prompting engineering and optimized model workflows. Currently fine-tuning and quantizing VLMs for efficient inference, with deployment on AWS Lambda, SageMaker and EC2 to ensure scalable, low-latency processing. Supports 13+ languages and multiple file formats, enabling domain-specific validation algorithms for high-quality field-level verification.