As AI auditing in hiring processes gains importance, New York City now requires companies to audit their AI systems for bias. However, the novelty of this requirement has led to several challenges, including loopholes and employee resistance, stemming largely from the autonomy companies have to self-audit and report results.

My research investigates the perspectives of lawmakers, managers, and AI system builders on AI auditing in HR. I aim to enhance trust in these systems by improving transparency, interpretability, and privacy by adding human feedback into the algorithms. This relatively unexplored area offers a significant opportunity for impactful research.

Currently discussing opportunities with several professors in the intersection of HCI, Business, and Law.

This project attempts to enhancing human task performance in performing clustering tasks by addressing the perceptual biases introduced by variations in scatterplot sizes. It started with my personal goal to reduce bias and help people create more accurate decision-making based on visual data interpretation.

To address these biases, I created a model to offset the influence of plot size on cluster recognition based on human perception experiments. Additionally, I developed a system that automatically adjusts scatterplot and point sizes to maintain consistent cluster granularity across different viewing conditions.

Advised by Jinwook Seo, we presented these findings for countering bias in visual data interpretation. The solutions proposed have been further detailed as a paper submission (VIS '24), where we outline how to maintain consistent data interpretation in different screen displays.

Question, Answer, and Connect began with personal frustrations regarding how Q&A massive group chats used in courses were extremely uncomfortable. Our initial low-fidelity demo, using Piazza-style bullet-type Q&A boards, revealed that more people preferred real time chats. To address the problem of asynchronous, untimely, and repetitive questions, we aimed to maintain the real time, convenient format of chatrooms while solving these issues.

By analyzing collected data, we used large LLMs to simulate student personas, predict frequently asked questions, and test if these questions were asked. I led the design and implementation of the system, conducting preliminary studies and low-fidelity demos, designing prototypes, and refining user experience through iterative testing.

Collaborated with KAIST KIXLAB's Hyunwoo, Yoonseo, and Naver's Conversational LLM search team "Cue" to create a taxonomy of conversational follow-up queries, discovering new user interactions from traditional LLM searches. We investigated how users' intent and method of asking the conversational agent correlated with user satisfaction.

During this research project, I was responsible for conducting thematic analysis for lab data, iterating and coding datasets to create the taxonomy, academic writing, and creating an automatic LLM taxonomy classifier that showed us user interactions we missed. This work was advised by Juho Kim submitted to the LLM4Eval workshop at SIGIR 2024.

Collaborated with David Lee from UCSC to help web dev novices learn while contributing to real-life systems. This system employs micro-role hierarchies to enable novices to make meaningful contributions to real-world projects. By dividing complex tasks into manageable micro-roles, we facilitate a learning process that allows novices to contribute to ongoing projects without having to go through the initial tutorials, which can often lead to boredom, fatigue, and loss of passion, as well as unnecessary time and overhead.

We developed the Causeway platform, where learners collaborate to build websites for non-profit organizations. In our proof-of-concept study, participants successfully created static websites for refugee resettlement agencies, demonstrating the effectiveness of micro-roles in providing practical and educational value.

As the project lead, I managed a diverse team of fifteen graduate and undergraduate students to expand on the previous work that received an honorable mention at CHI 2019. My responsibilities were to lead the development and integration of micro-role hierarchies within the Causeway platform, enhancing its capability to support web development education through structured, real-world contributions. To test this system for an actual user study, I had to conduct proto studies, interviews, and thematic analysis to explore how we can encourage novices to contribute while learning. This project not only highlighted the potential of micro-role hierarchies to transform educational approaches in technical fields but also reinforced the importance of practical experience in learning environments.

We introduce a novel dimensionality reduction technique designed to address the shortcomings of well-known methods like UMAP and t-SNE. UMATO, or Uniform Manifold Approximation with Two-phase Optimization, enhances accuracy, scalability, and stability in processing high-dimensional data. As the third author, my role involved evaluating the UMATO algorithm and contributing to the development of an interactive demonstration. This demo serves as both a practical and visual supplement to our academic paper, allowing users to explore various datasets and adjust parameters to see the effectiveness of UMATO in real time. Through this project, we introduced a new dimensionality reduction method, providing a more reliable tool for researchers and data scientists to uncover meaningful patterns in complex data.

FashionAlign is a new AI system to guide the way novice users search for fashion items by bridging the gap between abstract style descriptions and practical search capabilities. Utilizing natural language processing and advanced visualization techniques, the system transforms vague style inputs into a curated visual display of fashion items, significantly improving the search experience. This approach allows users to search using intuitive, everyday language, aligning textual descriptions with visual features to ensure that results closely match user queries. The system's use of UMAP (Uniform Manifold Approximation and Projection) for dimensionality reduction facilitates an organized and visually appealing presentation of search results. Designed with a focus on user-centric principles, FashionAlign has demonstrated improved match accuracy and query refinement in user studies, making it a valuable tool for simplifying the fashion search process.

JDnews, short for Joong Ding (a playful term for middle schoolers akin to "Zoomer"), was created to engage middle schoolers in reading newspapers. Collaborating with LG AI Discovery Lab, we created a system that automatically takes newspaper and generates summaries with vocabulary suited for middle schoolers, using data from the Korean traditional dictionary to train and create our own LLM. The project earned a runner-up award at the camp, allowing several students to attend a summer camp at Stanford. I led a group of four amazing students and mentored the implementation, design, and ideation process.

During my service as a social service agent, I contributed to the development of a critical alert service designed to protect civilians by notifying them of any nearby sex offenders. This service utilizes map data and information provided by the Ministry of Gender Equality and Family to deliver timely alerts. As a front-end designer and developer in the early stages of this project, I developed and designed the user interface and conducted demo user testing to ensure that users felt safer and protected with the program.

The SEIR model, a robust mathematical framework, is utilized to simulate various pandemics, from the spread of negativity on social media to the dynamics of COVID-19. This project applies the SEIR model to compare the progression of COVID-19 in South Korea and the U.S., exploring how different public health responses influenced disease dynamics. By modeling these scenarios, we can assess whether specific phenomena cascade through populations. The use of the SEIR model calculator enhances our analysis, providing a detailed comparison of the pandemic responses in both countries.

This project is a statistical analysis and data visualization exploration using the Datasaurus Dozen, a modern take on Anscombe's Quartet that demonstrates the limitations of relying solely on statistical summaries. In this project, I created visual representations with D3 to illustrate how different datasets can appear despite having the same statistical measures, highlighting the necessity of graphical analysis in visualization.

I developed an interactive visualization system using D3 and React to explore the effectiveness of various visual encodings in data representation. This system allows users to dynamically assign different data attributes—such as movie genres, ratings, and revenue—to visual channels like position, color, opacity, and size. By manipulating these visual properties through an intuitive interface, users can analyze how changes in marks and channels impact the clarity and efficiency of data presentation. The goal was to provide a hands-on tool for examining the expressiveness and effectiveness of visual idioms in real time, using a dataset of movie reviews to demonstrate how different visualization approaches can reveal or obscure patterns in the data. While implementing the system, I explored discrepancies in how different stimuli in visualization features are perceived, which inspired my first solo research topic.

I developed a visualization system to effectively represent high-dimensional data in a comprehensible two-dimensional format using the Star Coordinates algorithm. This system enables users to dynamically adjust axis vectors via mouse interactions, offering a live update of data projections that enhances understanding of complex datasets. Additionally, I implemented CheckViz, a component that visualizes distortions such as Missing and False Neighbors, using a color-coded Voronoi diagram to indicate the level of distortion for each data point. This approach not only provides insights into the data integrity after projection but also helps in assessing the trustworthiness of the multidimensional projection. By integrating these tools into a user-friendly interface with real time feedback and a clear legend for distortion interpretation, the project successfully simplifies the complexity of high-dimensional data analysis, making it accessible and actionable for users.

In this project, I utilized the Schelling model, a renowned agent-based simulation, to explore the movement and interaction patterns within multiracial communities. This model, which incorporates varying preferences for neighborhood racial compositions, helps to demonstrate how individual biases can lead to significant levels of residential segregation, even when those preferences are relatively mild. By adjusting parameters related to tolerance and neighborhood satisfaction thresholds, the simulation provides insights into how subtle changes in individual behaviors can influence the broader community structure. This study not only sheds light on the underlying mechanisms of segregation but also assists in understanding potential strategies for fostering more integrated communities.

As an intern at Samsung, we aimed at advance mobile interactivity within augmented reality. My role was to develop and benchmark InputOVR handtracking for Samsung's yet-to-be-released Vision AR goggles. This initiative involved the integration of handtracking cameras and passthrough functionalities, crucial for creating immersive and intuitive user interactions. I was responsible for back-end operations using vulkan api to include the implementation of the handtracking technology and testing phases to ensure performance.

In the project "Modeling Character Interactions Using Graph Theory Analysis," I applied graph theory—originally learned through analyzing social networks on Twitter—to explore the intricate character dynamics in "Les Misérables." Known for its complex and lengthy character interactions, this classic novel provided a rich field for examining how fictional relationships can be systematically mapped and understood. Employing graph theory metrics such as closeness, betweenness, and eigenvector centrality, I was able to identify key characters like Jean Valjean, Marius, and Gavroche, highlighting their crucial roles in connecting and influencing the story's development. This analysis delineated distinct communities within the narrative, such as revolutionary students and authority figures, showcasing how various character groups contribute to the novel's elaborate social dynamics. Through this comprehensive network analysis, I gained deep insights into the structural and thematic components of the narrative, significantly enhancing the understanding of character influence and community interactions within the text. This project demonstrates the utility of graph theory as a powerful tool for unraveling and understanding key interactions in any fictional work, opening avenues for deeper literary analysis and interpretation.

In this project, I implemented the PageRank algorithm using MapReduce to analyze character significance in fiction novels. By adapting a web page ranking technique to literary analysis, this method allows us to determine the importance of characters based on their interactions within the story. This approach not only highlights which characters are central to the narrative but also provides a scalable, efficient way to handle large datasets, making it suitable for examining complex novels with multiple characters and intricate relationships. This innovative application of PageRank and MapReduce opens new possibilities for deepening our understanding of character dynamics in literature.

"Nostalgia vs. Hype? Understanding User Sentiment in Video Games" was a project born out of a personal bet to determine which video game had a more significant social impact: the classic 'MapleStory' or the more contemporary 'Overwatch'. Conducted in 2016, this study involved analyzing a large corpus of Korean language data sourced from social media to conduct sentiment analysis. The objective was to assess whether nostalgia or hype more strongly influences user sentiment towards these games. This solo project not only fulfilled a personal curiosity but also contributed insights into the dynamics of player engagement and emotional responses within the gaming community.