Abdumannon H.

Computer Vision Engineer | Real-Time Video Analytics, Face Recognition, OCR & TensorRT/DeepStream Computer Vision Engineer with 50+ delivered projects covering the full pipeline: data cleaning and annotation, model training, optimization, and real-time deployment. I specialize in object detection, multi-object tracking, face recognition, re-identification, OCR and document AI. My production systems run on NVIDIA DeepStream, TensorRT and Triton Inference Server, for clients in retail, finance, healthcare, logistics, transportation and agriculture. I'm most interested in projects that take a model from prototype to a reliable production system: real-time video analytics, edge deployment, and large-scale document processing. Reach out to discuss your project. I'm happy to suggest an approach, point out risks early, and give a realistic estimate before we start. SELECTED PROJECTS. Real-Time CCTV Retail Analytics with Face Recognition Built a real-time video analytics system for retail stores running on 4 live RTSP cameras: 2 at entrances and 2 at exits. It handles automated attendance for employees and re-identifies returning customers to build a behavior profile for each visitor. I trained a custom face detector and an ArcFace-based recognition model myself, tuned to work on masked faces and real-world CCTV footage. Both models were compiled to TensorRT engines and served through NVIDIA Triton Inference Server, with a DeepStream pipeline processing all streams in parallel. Real-Time Workplace Analytics & Zone Monitoring (Lead CV & Deployment Engineer) Led the development of a system that measures how much time each person spends at their desk, from live CCTV feeds. Each camera covers 5 desk zones. I owned the full pipeline: data cleaning, custom annotation, and training YOLO detection and person re-identification models, so each person keeps the same ID and their desk time is logged accurately even when they leave and come back. Deployed with TensorRT, Triton and DeepStream for low-latency edge inference. Large-Scale Document AI: Bank Statement & Invoice Extraction Built a document processing pipeline that extracted structured data from 500,000+ bank statements and pulled key fields from invoices in many different layouts. For documents with a fixed structure, I fine-tuned PaddleOCR on custom-annotated data to improve recognition accuracy beyond the off-the-shelf model. License Plate Recognition. Detection and OCR pipeline for Korean and Kazakh license plates, including automatic plate replacement in images and video. Vehicle Tracking & Speed Estimation Combined YOLO detection with Deep SORT tracking to follow vehicles across frames and estimate their speed from traffic footage. Medical Imaging: COVID-19 & Viral Pneumonia Detection Chest X-ray classifier that separates COVID-19, viral pneumonia and normal cases with 95%+ accuracy. Also delivered: fruit and crop detection (apples, peaches, potatoes), and keypoint-based ID card and document localization feeding into OCR. TOOLS AND INFRASTRUCTURE Modeling & training: PyTorch, TensorFlow, Ultralytics YOLO (v5–v11) Face & re-identification: custom face detection, ArcFace, person ReID Tracking: Deep SORT, multi-object tracking OCR & document AI: PaddleOCR (fine-tuned), keypoint detection Deployment & inference: NVIDIA DeepStream, TensorRT, Triton Inference Server, ONNX Video: live RTSP streams, multi-camera pipelines Classical CV: OpenCV Language: Python Top Rated on Upwork. I write clean, documented code, give clear progress updates, and deliver on time.