Junbum Cha

KARA (KAkao brain Radiologic Assistant) is an AI-powered project that generates radiology reports from chest X-rays or CT scans, aiming to alleviate the shortage of radiologists. The KARA-CXR model, tailored for chest X-rays, exhibits performance competitive with junior radiologists and significantly surpasses general large multimodal models like GPT-4V in generating accurate and coherent reports.

Reliability is crucial in healthcare systems. This project enhances reliability by introducing explainability, which helps clarify the reasoning behind a model's predictions. More specifically, we focus on our medical reports generation model, which generates reports based on medical images. Our aim is to provide a visual representation that highlights the specific areas of the image upon which the generated report relies.

The goal of our team is to create a world-leading OCR engine. My main contribution lies in the development of the scene text recognition engine. By effectively tackling the out-of-distribution problem, our OCR system effectively addresses real-world scenarios. As a result, our model achieves unparalleled performance in Korean, Chinese, and Japanese, which are the primary languages aligned with our business objectives.

Korean fonts are composed of 11,172 letters, even not including other symbols, which makes creating a new one a labor-intensive and expensive task. This project has revolutionized the process by using deep generative models, specifically generative adversarial networks (GANs), to generate new fonts with almost 0.1% of the effort. This project opens the era of personalized Korean fonts where anyone can have their own font.

Handol is a project aiming at a superhuman-level go AI for business use. We've successfully reproduced AlphaGo, AlphaGo Zero, and AlphaZero, and our system achieved the runner-up place in the world go AI competition. We've also won several matches against professional players, including handicap matches. Using this system, we offered additional services such as AI odds prediction, AI hints, and AI matches on the company's Go game platform. We've also expanded the system to create automated level design for various turn-based games beyond go.

This project aims to cluster duplicate images within user-uploaded images on a cloud service. Specifically, a duplicate image is defined as any image that has been manipulated in some way, such as applying photo effect filters, adding stickers, resizing, or cropping. We utilized a SIFT-based image correspondence algorithm for image matching and an approximated kNN algorithm for search efficiency.

The motivation behind this project is to tackle the issue of overestimation resulting from cherry-picking in generative adversarial networks (GANs). In 2017, GANs were the most trending topic in deep learning, yet the absence of quantitative evaluation metrics led to a proliferation of diverse GAN algorithms. Each algorithm claimed superiority over others based on non-objective qualitative evaluations. To address this, this project offers open-source implementations of various GAN algorithms and provides an unbiased and objective comparison without cherry-picking.

I began my career as a software engineer. Within the first three years, I co-founded a start-up and developed 9 iOS applications. Out of these 9 applications, I contributed to 4 of them across all stages of productization, including initial ideation, product planning, business modeling, design, marketing, sales, and operation, beyond just application development.