Jeiyoon Park

I'm an AI Engineer at SOOP Co., Ltd.

My research interests are Video Understanding, Large Language Model (LLM), Role-playing Language Agents, and Reinforcement Learning.

Contact: k4ke@korea.ac.kr

Github  /  Google Scholar  /  LinkedIn  /  YouTube  /  Instagram  / 

[Nov 2024] I showcased SOOPI (SOOP Personalized Intelligence), a video assistant based on a streamer persona, at G-STAR 2024.
[Mar 2024] I will serve as a program committee member at BEA-NAACL 2024.
[June 2022] I came in first place at 2022 Microsoft Azure Virtual Hackathon.
[Mar 2021] I interned as a NLP researcher at Naver Webtoon (Webtoon AI), where I had a wonderful time working with Chaeheon Gwak and Seho Kim, and being mentored by Daesik Kim also.

Azure Virtual Hackathon aims to solve various social problems we face, such as climate change and AI for accessibility, by using Azure, Microsoft's AI solution. This work introduces "Hi Azure!: Helping video creators who are visually-impaired edit videos using Azure Cloud AI!", a framework for increasing accessibility to video editing and content creation for visually impaired creators by extracting visual features from video, converting them to voice, and providing them as frame-level descriptions. We came in first place at this Hackathon and additionally I got first place at Microsoft Cloud Skills Challenge (1/63 = 1.58%).

We postulate that the frames where emotional information appears frequently in a video are correlated with highlights. So our model first extract visual features and audio features from a video simultaneously. And then we employ attention mechanism with these features. Finally our model predicts importance scores from the Emotion-attended features and generates highlights based on the scores.

In this project, we propose a novel meta-reinforced multi-domain task-oriented dialog agent, which exploits autonomous-growing knowledge base and extends the existing single-domain agent algorithm to multi-domain. Our method is expected to handle not only inferring about information that is not shown in the training process or the relations between domain, slot, and value triples, but also responding to domains other than the existing learning domain.

The 31st Korean and Korean Information Processing Conference (HCLT 2019) was held at Daejeon Korea Institute of Science and Technology (KAIST). We were awarded the Encouragement Award at the 2019 Korean Information Processing System Competition.

The ITRC Global Makerton 2019 was held at Hanyang University. Our team proposed a simple noise canceling software for target noises (e.g. noise from cafe or factory etc.) and won second place.

We proposed a simple closet that help people to keep their clothes clean and efficient. Also, we invented a novel hanger sensor and actuator for this project. I collaborated with the designer "Jeongse" and we were awarded the runner-up prize.

We proposed "smart hospital scenario using future robot technology". I collaborated with the nurse "Minji", designer "Yeonkyung", and computer scientist "Kihwan" for this project. We won the Encouragement award.

This challenge consists of four missions: (1) Pattern recognition for the trafic sign. (2) Autonomous parking. (3) Stop drving in front of a pedestrian crossing. (4) Any missions that you want to perform. We won the bronze medal in this challenge.

We proposed a novel racing simulator which is fun and safe to enjoy at a much lower price than previous racing simulators. To this end, we invented new elastic actuator. We won second place.