Hyogeun Oh (오효근)

Tmax WAPL

• Developed backend infrastructure for WAPL, a collaboration platform, using microservices architecture (MSA) within an 11-member team.

• Developed scheduling functionality utilizing a Netty-based in-house Java backend framework, supporting the efficient creation, deletion, and retrieval of schedules.

AgileSoDA

• Managed the entire lifecycle of machine learning services (annotation, modeling, training, deployment) and oversaw Kubernetes-based IDC infrastructure within an 11-member team.

• Researched and developed models for text, signature, and checkbox detection, as well as information extraction, for the AI optical character recognition (OCR) product TwinReader.

• Developed a Python backend for model serving and optimized pipelines to enhance efficiency.

• Streamlined the development process by creating a solution that resolved versioning challenges and reduced build times through the separation of backend dependencies.

• Executed AI projects and proof of concept (PoC) to meet client specifications.

SiM Lab. (Smart intelligent Manufacturing system Laboratory)

• Researched and developed diagnostic models utilizing time-series and vision data for real-time monitoring of process conditions in industrial-scale manufacturing systems.

• Published two SCI(E) Q1-ranked research papers on feature selection algorithms, advancing efficient machine learning methodologies in industrial applications.

• Executed 10+ government and industry-funded projects, further enhancing research capabilities in machine learning.

SiM Lab. (Smart intelligent Manufacturing system Laboratory)

• Developed bearing condition diagnostic model and graphic user interface.

MRV Lab. (Medical Robotics and Virtual Reality Laboratory)

• Developed virtual reality environments based on C++ and Unreal Engine.

Development and CI/CD Pipeline Construction of Python Library

• Zerohertz/zerohertzLib (GitHub)
• Python Library (PyPI)
• Docs (Sphinx)

• To reduce time consumption and inefficiency from reimplementing commonly used functions, developed and published a custom Python library on PyPI and GitHub Releases to enhance efficiency and code reusability across projects.

  

• Built a GitHub Actions-based CI/CD pipeline (migrated from Jenkins) to automate repetitive tasks such as formatting, unit testing, and deployment, streamlining the process for feature additions and bug fixes.

  

• To prevent unnecessary deployments from non-production changes like documentation updates, implemented a detailed branching strategy on GitHub and set up dedicated pipelines for code segregation.

  

• Simplified version tracking by building a pipeline using the GitHub API to automatically generate and publish release notes to GitHub Releases, improving transparency across development cycles.

• Ensured easy access to comprehensive project guidelines and function usage by creating a Sphinx-based documentation pipeline, deploying it via GitHub Pages for consistent and up-to-date project documentation.

  

Python Library for Pre/Post-Processing, Visualization, and Model Backend

• Packaged frequently used classes and functions within the model backend into a Python library to streamline development processes.

• Utilized Docstring to document functions and classes, enhancing code clarity and team collaboration, while maintaining library integrity through type hints and PyTest.

• Faced with significant compatibility issues due to inconsistent libraries and formats for model outputs, standardized the data format for preprocessing and model inference visualization, enabling consistent visualization and resolving unexpected compatibility problems.

• Addressed inefficiencies in post-processing due to Python-native functions with high time complexity by optimizing them with Cython-native functions and improving time complexity. (inference time decreased by 74.12%)

  

• Developed a unified class and inheritance structure for Triton Inference Server.

AI-based OCR Solution, TwinReader

• Developed models for document area detection, rotated document classification, and detection of text, signatures, and checkboxes, along with a Python backend for model deployment.

• Integrated project source codes using Git Submodule to facilitate a smooth collaborative environment.

• Implemented a pipeline using GitLab CI and Kubernetes to separate backend dependencies from code and weights, deploying through the Kubernetes sidecar pattern, which resolved versioning challenges and significantly reduced build times for the model backend.

  

• Faced with excessive GPU usage during model deployment, resolved the issue by identifying and fixing a memory leak through GPU resource monitoring and logging. (GPU memory usage reduced by 47.9%)

  

• Reduced inference time for a text detection model, where frequent calls made optimization critical, by utilizing TensorRT-based quantization. (inference time decreased by 87.31%)

  

• Encountered low accuracy in document rotation classification, addressed by performing batch inference on image tensors rotated in four directions and averaging the results. (improved accuracy by 2.01%p)

  

Information Extraction Model Development for Trade Document Processing

• Performed clustering, annotation, preprocessing, training, and deployment to develop a model for extracting information from a wide variety of trade document formats.

• Faced with the challenge of categorizing large volumes of unstructured PDF documents, developed an AI OCR-based pipeline utilizing OCR results and LLM prompting to efficiently classify and sort documents. (achieved 93.75% accuracy)

• To address the high time and cost demands of large-scale data annotation requiring expert knowledge, accelerated the process by implementing pre-labeling through an ML backend using Label Studio SDK, significantly reduced annotation time and costs.

• Encountered difficulties in manually checking complex human errors during annotation review, developed a Streamlit-based GUI to allow easy detection and correction of these errors through simple configurations.

AI-based Automotive Damage and Depth Recognition Product, TwinCar

• Developed a vehicle type classification model for PoC execution.

• Conducted research and development models for filter, part recognition, repair type, and damage type, along with the model inference pipeline.

• Developed a demo page using Streamlit and deployed it on Kubernetes.

AI Diagnostic Service for Burn Patients

• Developed models for segmentation of burn areas and severity diagnosis in burn patients.

• Designed and developed an API for model deployment using Triton Inference Server.

  

On-premise Kubernetes Cluster

• Zerohertz/k8s-on-premise (GitHub)
  
• Blog posts
  

• Installed Kubernetes using Kubeadm on an on-premise environment to enhance understanding of Kubernetes architecture and practical usage.

• Secured deployed services by implementing HTTPS protocol and Google OAuth2 through Traefik.

• Established GitOps by automating build and deployment processes using GitHub Actions and Argo CD.

• Built a node status monitoring GUI leveraging Node Exporter, Prometheus, and Grafana.

• Automated various tasks using Apache Airflow integrated with KubernetesPodOperator.

• Set up a Docker image build and deployment pipeline using Jenkins and Kaniko.

Mosaic Classification

• Team-BoonMoSa (GitHub)
  

• Developed a YOLOv5 based logo segmentation model.

• Constructed a model deployment server on Amazon EC2 Inf1.

  

• [Patent] Apparatus and method for predicting coating layer thickness in slot-die coating process

• Inventor: Changwoo Lee, Minho Jo, Yoonjae Lee, Hyogeun Oh

• Application number: 1020220017419 (2022.02.10)

• Grant number: 1025842600000 (2023.09.25)

• [SCI(E)] Classification and Redundancy Quantitative Evaluation Algorithm for Highly Efficient Fault Diagnosis of Rotary Machines in Roll-to-Roll systems

• Author: Hyogeun Oh, Jaehyun Noh, Changbeom Joo, Gyoujin Cho, Jeongdai Jo, Changwoo Lee

• Journal: Measurement [Impact Factor 5.60, JCR Top 17.19%]

• [SCI(E)] Feature Selection Algorithm Based on Density and Distance for Fault Diagnosis Applied to a Roll-to-Roll Manufacturing System

• Author: Hyogeun Oh, Yoonjae Lee, Jongsu Lee, Changbeom Joo, Changwoo Lee

• Journal: Journal of Computational Design and Engineering [Impact Factor 6.16, JCR Top 10.87%]

• Customized a technical blog based on the Hexo NexT theme to document and share solutions to challenges encountered during personal learning and professional work.

• Achieved 1,500 MAU and 2,600 monthly page views by consistently writing 200+ posts since 2018.

• Implemented an automated data ingestion and preprocessing pipeline using GitHub Actions to enhance data workflow efficiency.

• Delivered insights to technical research personnel (전문연구요원) through data visualizations created with Matplotlib, supporting decision-making from multiple analytical perspectives.

• Implemented an automated data ingestion and preprocessing pipeline using GitHub Actions to enhance data workflow efficiency.

• Delivered insights to skilled industrial personnel (산업기능요원) through data visualizations created with Matplotlib, supporting decision-making from multiple analytical perspectives.

• Identified and verified a dependency mismatch with Image.Resampling.BILINEAR (Pillow >=9.1.0).

• Conducted version testing and recommended an update to Streamlit’s requirements, improving reliability for developers by preventing compatibility issues.

• Enhanced environment compatibility by removing python==3.7 from requirements.txt, enabling broader setup compatibility for SPTSv2.

• Corrected variable type mismatch by aligning the depths variable to a list type for consistency with its default value, enhancing code clarity and reducing runtime errors.

• Generalized configuration by implementing customizable parameters like max_length in data loading and model setup, improving SPTSv2 adaptability for varied use cases.

• Optimized memory usage in inference by adding the @torch.no_grad decorator in predict.py, significantly reducing GPU memory requirements.

• Resolved IndexError during training with customized data by fixing shape mismatches in GT data, ensuring stability in data handling.

• Addressed tensor dimension errors and generalized prediction, evaluation, and visualization processes.

• Added a motion.duration parameter in Hexo's NexT theme _config.yml, enabling flexible configuration of motion animation duration.

• Modified source/js/motion.js to retrieve motion.duration dynamically, with a default value fallback for robustness.

• Contributed minor wording refinements to improve grammatical accuracy.

Konkuk University, Seoul, Korea

• Thesis: A study on feature variable density for a highly efficient diagnosis model of rotary machine in roll-to-roll continuous process

• Advisor: Changwoo Lee

• GPA: 4.15 / 4.5

Konkuk University, Seoul, Korea

CKA (Certified Kubernetes Administrator)

전문연구요원

BOAZ 19th (Big Data Club)

• Study & Project (GitHub)