Roubish Kumar Pushkar

Building deep learning models for images, videos, and text data. Designing and building impactful Deep-Learning solutions for business use-cases. Actively researching about the latest AI and DL techniques.

• Code Development: Designed and implemented multiple scalable computer vision pipelines deployed across real-time surveillance and industrial use cases, improving system processing speed by up to 40%.
• Model Training Optimization: Trained and fine-tuned custom object detection models (CNNs, YOLOv5/v8, SSD) achieving detection accuracies of 95%+ in complex environments (night-time, occlusion, crowd).
• Software Maintenance: Maintained and optimized multiple production-grade vision applications, reducing false-positive alerts by 30% and increasing system uptime to 99.5%.
• Hardware Debugging: Diagnosed and optimized 100+ edge devices (NVIDIA Jetson Nano, Xavier NX, AGX Orin), improving memory efficiency and reducing power consumption by 20–25% per device.
• PoC Development: Developed proof-of-concepts of which some were adapted into full-scale deployments — including ANPR, intrusion detection, and industrial safety systems.

Python, TensorFlow, PyTorch, OCR, CNNs, ANN, Object Detection, Semantic Segmentation, Image Classifi- cation, Deep Learning, Software Development, NVIDIA Jetson, Edge AI, DeepStream SDK, OpenCV, YOLO, SSD, Computer Vision, MySQL/NoSQL, GCP, GCF, AWS, FFmpeg, GitHub, Docker.

• Designed and developed an end-to-end ANPR system by integrating PaddleOCR with DeepStream for real-time license plate detection.
• Optimized accuracy and reduced processing time, ensuring efficient recognition in various lighting and weather conditions.
• Conducted comprehensive testing and debugging for seamless deployment in real-world environments.

• Implemented video analytics using DeepStream Python to measure customer waiting and service times at retail stores.
• Provided data-driven insights to optimize workflow efficiency and enhance customer experience.
• Designed a scalable and automated monitoring system for continuous improvement in service management.

• Developed custom AI models to monitor warehouse operations, including shutter status tracking and sack counting for logistics efficiency.
• Implemented vehicle entry monitoring, person intrusion detection (especially at night), and guard attendance tracking to enhance warehouse security.
• Integrated real-time alerts and visual dashboards to provide actionable insights for warehouse management.

• Designed AI-powered safety solutions to detect helmet and vest violations, hand glove violations, forklift overspeeding, and illegal vehicle parking.
• fire and smoke detection algorithms for early hazard identification and prevention.
• edge-based models on NVIDIA Jetson devices for real-time processing, ensuring workplace com- pliance and accident prevention.

Padecco India secured a contract for Mumbai Metro’s surveillance project, covering 80 metro stations with 15–20 cameras per station (FOB, Platform, Ticketing Area, Entry-Exit). The project involves processing 7 days of video data per camera to generate analytical reports for the Padecco team.

Use Cases Implementation:

• Male-Female Detection – Deployed custom YOLOv8 model for gender classification on FOB cameras.
• Head Count Detection – Used custom YOLOv8 model for real-time crowd estimation on platform cameras.
• Queue Length Estimation – Tracked queue lengths at ticketing areas to analyze congestion.
• Peak Hour ROI Analysis – Analyzed crowd density trends, peak-hour statistics, and male-to-female ratio in different regions of interest (ROIs) over time.
• Optimized Processing Pipeline – Integrated DeepStream Python pipeline for high-speed inference and efficiency.
• Hardware Acceleration – Used NVIDIA A6000, RTX 4090, and RTX 3090 for scalable batch inference.
• Automated Workflow – Developed Python Shell scripts for complete automation of video processing.
• Data Storage Post-Processing – Stored 7 days of inference results in Excel sheets, followed by analytical reporting.
• Video Format Standardization – Used FFmpeg MEncoder to convert .DAT, .AVI, .H264 to MP4 for seamless processing.
• Team Collaboration – Worked in an 8-member tech team, supported by 2 additional team members.

Technologies Used: DeepStream 6.1.1/6.2, YOLOv5 CrowdHuman model, YOLOv8s Male/Female model, OpenCV, Linux Scripting, GStreamer, Pandas, NVIDIA.