Project Gallery

Students worked on a semester-long design project, in which they designed and built an AI-infused interactive application to explore various aspects of human-AI interaction. This poses interesting and unique design challenges in terms of the mental model of AI, explainability, user control, feedback, visualization, etc.

Ten student teams came up with creative designs, and this page gives a snippet of their design process.


Members: Aitolkyn Baigutanova, Gabriel Lima, Assem Zhunis

Taking notes, although helpful in revising lectures, might be burdensome as it might be difficult to follow lectures at the same time. An interactive platform for video lecture summarization, where users can obtain an automatically generated summary from YouTube video subtitles and edit them as they see fit. We use a BERT-based extractive summarization model for generating editable summaries from video transcriptions, where users can navigate through video lectures by clicking on specific parts of the text, and also search for more information about specific terms.

Check out the LIVE INTERFACE. Google Chrome recommended.


Members: Ayan Mukanov, Timothy Manoj Mathews, Naziya Zhakupova

Traditional word clouds are not so comprehensible because ideally they should be subdivided into semantic clusters that are visually distinguished (by spacing and color). We combine research on semantic word cloud generation based on word embeddings, with research on the optimally comprehensible appearance, to let users make ideal word clouds easily. Additionally, for the first time in this area, we propose and implement a "computer suggests & human decides" approach.

Check out the LIVE INTERFACE.


Members: Jinwoo Kim, Jungwon Lee, Seonghoon Jeong

A black-box inpainting AI can ignore user intention without explanation, harming the user experience. To solve the problem, we injected a backdoor to a trained black-box AI and connected it to an intuitive coloring UI. Our approach allows a joint understanding & control of the inpainting process, and is fully unsupervised, unlike most of the interactive AI.

Check out the LIVE INTERFACE.

Fantastic 4

Members: Farid Talibli, Mahmoud Asem Mohamed Abdulaziz, Beomki Kim

There are various issues on facial expression data about privacy, bias, trustworthiness. Therefore, to resolve these issues, we propose a novel crowdsourcing platform to gather facial expression data. This platform is a game to reduce the boredom of gathering information, and stores only facial landmarks to secure user privacy and mitigate bias (skin color , ethnic features,gender) results from training on facial data .


Members: Soyoung Yang, Youngjae Jang, Jihyeong Hong

We aim to solve the problem of relationship status analysis using chat logs done by AI/template based method not being open to receiving feedback. To solve this issue, we allow users to first observe the sentimental analysis result, report their own thought on relationship, and then see the cooperated decision on the relationship. Our approach is unique because it uses objective analysis of AI while considering human's thought which comes from details between lines.

Check out the LIVE INTERFACE.

SHY team

Members: Shyngys Aitkazinov, Hyunsung Cho, Yewon Kim

Despite the variety of existing AI pronunciation practicing models, there is no AI-interactive platform for non-native English speakers to evaluate and practice English pronunciation skills with __personalized scripts__ . To solve this problem, we propose Scooby, a AI powered platform where users input their scripts, practice the pronunciation and receive the feedback from AI to further improve the English speech skills. Unique features of Scooby include ii) enabling a user's __personal text input__, ii) visualizing the speech-to-text results with __wrongly spoken parts colored__, iii) providing sophisticated __phonetic-level analysis__ from AI, iv) __scoring the user speeches__.

Check out the LIVE INTERFACE.

The Sorting Hat

Members: Cholmin Kang, Zee Wung Shin, Jaeryoung Ka

Many students and instructors leverage auto-translators for studying, however, they often lack domain-specific knowledge which often leads to wrong terminologies and hinder the learning progress. Existing translation tools used to rely on community systems to fix awkward expressions, but such community systems showed shortcomings coming from them being not primarily designed to provide domain-specific word correction. In this work, we propose "The Sorting Hat" that leverages the power of users with a diverse level of domain expertise and lets them directly interact with the model to correct wrong terminologies in educational context in a fast and reliable way to overcome such issues.

Check out the LIVE INTERFACE.


Members: Junyong Park, Dinmukhamed Mailibay, Wasachon Chaisirirat

It is difficult for people *shadowing* with movies (copying what characters in the movie say) to get a sense of how well they are doing or how they can improve. We developed a website where people can record themselves shadowing dialogues in the movie and then get a score on how similar they are to the character in the movie. In order to make this interface more user-friendly, we focus on making this pronunciation evaluation process configurable by allowing the user to adjust what they are scored on, as well as explainable by providing a simple explanation for the scores.

Check out the LIVE INTERFACE.


Members: Le Viet Anh, Yourim Shin, Duc Khanh Thi Nguyen

Help students apply to KAIST find the answers for their questions efficiently and save their time. Our solution is 24/7 chatbot. We use NLP models to tackle the issue.

Check out the LIVE INTERFACE


Members: Quinones Robles, Willmer Rafell, Sakonporn Noree, Yoonhoi Jeon

Our system, YOYO, allows you to do yoga on your own, wherever you want, whenever you want, without having to go to the gym, and without expensive private tutors. All you need is a space for a yoga mat and a device with a camera (cell phone, laptop). Our system, using Deep learning for Computer vision, extracts the skeleton from the instructor and user's postures, compares them and displays them on the screen. In addition, the user can directly adjust the sensitivity of the model, and the system provides feedback to help users make better use of the system.

Check out the LIVE INTERFACE.