Saravanan H

Solution architect specializing in video analytics and AI systems with over 8+ years of experience in designing scalable computer vision solutions for edge and cloud environments. Expertise in developing end-to-end AI pipelines utilizing deepstream , openCv, PyTorch, and OCR systems, with a focus on system architecture and performance optimization. Proven track record in designing distributed architectures, optimizing compute and memory usage, and delivering high-accuracy models while effectively managing false positives and edge cases. Skilled in integrating jetson edge devices, camera pipelines, and cloud platforms.

Loss Prevention System (Retail Video Analytics using DeepStream & C++)

Designed and implemented a real-time AI-based surveillance system to detect and prevent shoplifting in self-checkout kiosks and cashier counters.

Tech Stack: NVIDIA DeepStream, C++, GStreamer, OpenCV, PyTorch, WebSocket, SQLite

Key Contributions:

• Architected an end-to-end video analytics pipeline using NVIDIA DeepStream SDK (C++) for high-performance inference
• Integrated YOLO-based object detection models for identifying persons and products
• Implemented multi-object tracking (DeepSORT / NvDCF) for tracking customer behavior across frames
• Designed logic for theft detection scenarios:Item picked but not scanned
  Suspicious motion near checkout
  Object disappearance from ROI
• Optimized pipeline using GPU acceleration to achieve real-time performance (low latency, high FPS)
• Reduced false positives/negatives through threshold tuning and edge-case handling
• Built event-driven alert system using WebSocket for real-time notifications
• Designed backend APIs for incident reporting and audit logs

Architecture Highlights:

• Edge-based inference using GPU (DeepStream)
• Cloud/API layer for reporting and analytics
• Modular pipeline enabling scalability across multiple camera streams

• Design end-to-end architecture for video analytics systems
• Define edge, GPU, and cloud processing layers
• Optimize pipelines for performance and scalability
• Build ML pipelines and improve model accuracy
• Handle false positives/negatives and edge cases