Abhay Kumar Tiwari

Computer Vision Engineer with hands-on experience building scalable detection, recognition, and segmentation systems, deployed end-to-end on NVIDIA Jetson and embedded edge devices. Specialized in real-time YOLO-based pipelines with deep expertise in OpenCV, Transformers (ViT/UNETR/VLM), TensorRT optimization, and ONNX-based model deployment. Proven ability to deliver low-latency, production-grade vision systems within strict edge constraints of memory, latency, and throughput.

Published researcher in transformer-based medical image segmentation (Springer). Strong in end-to-end AI system development — from model architecture design and optimization to deployment and CI/CD automation.

Computer Vision Engineer at eSec Security Consultants Pvt. Ltd. (2025-08 – Present)

• Architected a modular NVIDIA DeepStream pipeline for real-time inference across 100+ concurrent RTSP streams, powering four production applications: chain-snatching detection, violence/action recognition, traffic congestion analysis, and weapon detection.
• Independently designed and built each application (YOLOv7-Pose, YOLOv8n, YOLOv11-Pose, PoseC3D) on a shared tech stack, containerizing each with Docker for isolated, independently scalable deployment.
• Prototyped on NVIDIA Jetson for edge validation, then re-engineered and deployed on NVIDIA L4 GPU servers for scaled inference, using TensorRT and ONNX for cross-platform optimization.
• Built a FastAPI control layer for remote Docker start/stop and live JSON-based configuration — RTSP updates, per-app stream limits, tunable inference parameters — exposed via a UI, with MQTT-based real-time alerting.
• Built the end-to-end flow — detection, embedding extraction, vector indexing, and real-time match/alert — achieving 87ms matching latency at 96.8% accuracy in production.
• Deployed on NVIDIA Jetson (Orin/Nano) for real-time edge inference, optimized via TensorRT.
• Deployed dual production pipelines on Axis Camera (ACAP) edge hardware — PPE compliance detection and fire/smoke detection — using Axis-native APIs (Larod, VDO Capture) for on-device DLPU inference.
• Optimized a custom YOLO model via TAO Toolkit and quantization, reducing size 75% (42.2MB ONNX FP32 → 10.5MB TFLite INT8) while retaining 97.5% accuracy at 30 FPS.
• Implemented OC-SORT tracking for robust multi-object tracking on constrained edge compute.

Research Assistant at National Institute of Technology, Durgapur (2024-05 – 2025-05)

• Conducted research on transformer-based 3D medical image segmentation under faculty supervision, resulting in a publication in prestigious Springer (2026).
• Designed and implemented a novel UKANETR architecture integrating KAN based MLPs and Squeeze-and-Excitation blocks, achieving Dice scores of 0.86 (Whole Tumor), 0.72 (Tumor Core), and 0.70 (Enhancing Tumor) on BraTS 2020, surpassing the SOTA UNETR model.

M. Tech in Computer Science and Engineering – National Institute of Technology, Durgapur (NIT DGP) (2023 – 2025)

B. Tech in Computer Science and Engineering – Aliah University, Kolkata (2016 – 2020)