Jungho Cho Email & Phone Number

Yongin-Si, Kr

대한민국 용인

Advanced Large Language Model (LLM) Development - Developed Solar Pro Preview and Solar Pro, advanced LLMs tailored for structured text understanding tasks, which outperformed most single-GPU models in general benchmarks and set a new standard for high-performance LLMs. - Created synthetic datasets with over 3 million samples for structured text understanding and over 2 million samples for code-related tasks. - Designed a unified evaluation pipeline for benchmarking LLMs across various document-related tasks like DocVQA and WikiTableQuestions.

대한민국 용인

Large Multi-modal Model (LMM) Research - Conducted research on LMMs for document understanding tasks such as information extraction, visual question answering, and visual grounding. - Developed a template-based dataset generation pipeline, producing over 40 million image-text pairs for pre-training and over 120 million pairs for instruction-tuning, significantly enhancing model capabilities.

대한민국 용인

Multi-modal Model for Information Extraction from Document Images - Developed an information extraction model using Swin transformer encoder and MBart decoder, improving image reading and context understanding through carefully designed pre-training tasks. - Created a comprehensive dataset with a crowdsourcing tool for labeling textual information, significantly improving model performance and enhancing user satisfaction. - The developed method became the core technology of the company's current information extraction product, establishing a strong foundation for its success and adoption in production environments.

대한민국 용인

Tagging-based Information Extraction from Document Images - Developed a BIO-tagging-based model for extracting information from document images and OCR text, achieving high F1 scores that secured multiple contracts. - Defined a company-standard evaluation metric and introduced a confidence score as a key KPI for evaluating model performance.Participation in ICDAR 2023 VQAonBD Challenge - Focused on answering questions based on business document images using tailored pre-training tasks with the open-source FinTabNet dataset. - Achieved a final score of 0.959, winning 1st prize in the competition.

Redwood City, California, United States

- Conducted research on NLIDB and developed a deep-learning-based text-to-SQL model using a graph neural network.- Surveyed previous deep-learning models in this field, analyzing their pros and cons.

대한민국 서울 구로구

- Developed a deep-learning-based AI bot for the game "Magic: ManaStrike" using the PPO algorithm.- Secured 2nd prize among all internship AI bots.

대한민국 경상북도 포항

- Developed a human detection model based on Faster R-CNN and a rule-based algorithm for human tracking.

M.S., Creative It Engineering

Activities and Societies: Database & Datamining Laboratory

Bachelor'S Degree, Creative It Engineering

Activities and Societies: Ctrl-D

Physics

Activities and Societies: Broadcasting Club