jakhon37 Goto Github PK

GitHub 🐱‍💻 | Google Scholar 📚 | LinkedIn 🔗

Machine Learning Engineer with a rich background spanning industrial and academic settings, committed to advancing artificial intelligence. Expertise includes supervised, self-supervised, and transfer learning techniques, with specialized skills in object detection, recognition, segmentation, and tracking. Excels in the development of 3D avatar creation algorithms using advanced morphable models. Proficient with AWS and edge computing devices for deploying scalable AI models that enhance IoT and virtual reality technologies, aiming to transform industry practices and elevate user experiences.

• Phone: 010-8025-0595
• Email: [email protected]
• LinkedIn: Jakhongir Nodirov
• GitHub: jakhon37
• Website: jakhon37.github.io

• Uzbek: Native
• English: Proficient
• Korean: Intermediate

• Programming Languages: Python, C/C++
• ML Frameworks: PyTorch, PyTorch3D, TensorFlow, Scikit-Learn, Numpy
• MLOps Tools: Docker, Docker-compose, Flask, FastAPI, Gradio, Streamlit
• Databases: MySQL, SQLite, Weaviate
• Development Tools: Git/GitHub, Docker, CI/CD
• Cloud Platforms: AWS EC2/S3, ORACLE Compute/Storage
• IoT & Embedded Devices: ROS2, Yocto, ONNX, TensorRT, Jetson Nano, Raspberry Pi
• Main Competencies: Object Detection/Recognition/Segmentation, Object Tracking, Action Recognition, OCR, Medical Imaging, Image Restoration & Enhancement, DeepFakes, Generative Models, Vision-Language Models, Large Language Models, Building End-to-End Pipelines, Deployment Pipelines

• Gachon University
  • Degree: Master’s degree in Computer Engineering
  • Period: 2020.09.01 – 2022.08.25
  • GPA: 4.17/4.5

• "3D Volume Reconstruction from MRI Slices based on VTK." In 2021 International Conference on Information and Communication Technology Convergence (ICTC), pp. 689-692. IEEE, 2021.
• “Attention 3D U-Net with Multiple Skip Connections for Segmentation of Brain Tumor Images” in Sensors 22, no. 17 (2022): 6501.

• Role: AI Software Engineer
• Period: 2023.02.01 – Present
• Responsibilities:
  • Research and Development: Continuously explore new deep learning techniques and frameworks.
  • Collaboration: Work with cross-functional teams to integrate AI solutions into projects and products.
  • Optimization: Enhance model performance and efficiency, including reducing latency and improving accuracy.
  • Deployment: Manage applications in production using Docker on AWS and other platforms.
  • Documentation: Create detailed documentation and reports on developed algorithms and models.
  • Maintenance: Regularly update models with the latest AI advancements.

• Techniques: Developed an IP Protection Application for logos, animation characters, and celebrity image verification.
• Technologies: Utilized Weaviate Database for storing intellectual properties and Fast Vector Search; Wide-ResNet for logo and animation body feature vector extraction; Arcface for human and animation face feature vector extraction; Streamlit for creating a responsive demo image vector search application.
• Innovation: Fine-tuned models on specific object patterns feature vector extraction and enabled protection of digital properties used on AIVAR’s T4U web application.

• Techniques: Pioneered a method for converting 2D images into 3D avatars (face and body) with detailed textures.
• Technologies: Leveraged FLAME, DECA, Albedo for 3D face mesh creation with high-quality texture; SMPL, HMR, BEDLAM for accurate 3D body mesh measurement and creation; Stable Diffusion for face image attributes correction and enhancement; Yolov8 for custom hairstyle classification model; CAP-VSTNet for face and body texture matching; Facer for comprehensive face parsing; Mediapipe and Dlib for real-time facial feature tracking and alignment; Scipy with SPIGA for FLAME prediction optimization; AWS C2, S3 for deployment with scalable compute and storage capabilities; Docker for ensuring consistent environments across development and production.
• Innovation: Enabled realistic 3D human avatar creation with lifelike textures and accurate body measurements, facilitating advancements in virtual presence technology.

• Techniques: Developed an application for manipulating facial attributes to customize face features in images.
• Technologies: Implemented face parsing to identify specific facial regions, landmark detection for precise area evaluation, and Stable Diffusion for advanced inpainting; employed OpenCV and Pillow for refined control over face image mask gradients.
• Innovation: Customized the Stable Diffusion model specifically for facial datasets, enabling dynamic modifications of face attributes such as eyeglass removal or altering mouth expressions.

• Techniques: Automated the process of license plate detection and recognition for Korean vehicles, including the preparation of training data through auto-labeling.
• Technologies: Utilized the Yolov and Darknet frameworks for real-time processing and OpenCV for image analytics; developed an auto-labeling system to efficiently prepare accurate training datasets.
• Efficiency: Achieved a streamlined identification process, significantly reducing the time required for vehicle registration and monitoring, and enhanced the accuracy of the machine learning model through improved training data quality.

• Techniques: Conducted comprehensive data preparation and architecture fine-tuning for road condition analysis.
• Technologies: Employed Detectron2 for precise segmentation and utilized OpenCV to analyze the quality of the segmented areas, with a focus on luminance and brightness metrics.
• Impact: Contributed to road safety improvements by enabling precise segmentation and quality assessment of road marks and fences, leading to better maintenance and hazard prevention.

• Techniques: Developed and fine-tuned a segmentation model for brain tumors in 3D MRI scans.
• Technologies: Employed PyTorch, Nifty for medical imaging, and custom 3D UNet for model training.
• Outcome: Achieved higher precision in tumor boundary detection, improving diagnostic accuracy.