web application, face-recognition

Due to the COVID-19 pandemic, online class increases and it is hard for both educators and students to actively interact with each other in class. Specifically, educators are having trouble on getting their students to participate in online class. In other words, it is difficult for educators to determine how much students are concentrating in online classes.

This application tries to capture the students' concentration on online class in real time. By providing the score of concentration, it helps educators gauge what students understand and what students may need educators to review again. In addition, it prompts students to participate in class more by trying to get higher score than their classmates.

android application, chat service

Since our school is far from the nearby subway station, students have hard time to get to school after arriving at the station. There are three options they can choose: a walk, a shuttle bus, or a taxi. Even though taking a taxi is expensive, some students decide to take a taxi when they are in urgency, such as being late for class. In response, our service starts with the question, "How about matching students who want to take a taxi from the station to school?"

This application provides a chat service for users who want to take a taxi to get to school but want to save their money by a shared ride. Up to four of the nearby users will enter one chat room, and then our application offers a simple Dutch Pay service assuming that one student has paid for a taxi.

android application, transportation service

The most commonly used public transportation for commuting in South Korea is by far "the subway". When taking the subway, passengers go back and forth to find empty seats. However, it is difficult to find empty seats, and even if it is found, it is often taken away by other passengers.

Here, our application provides a service for sharing boarding information of passengers on the same train. Using the public api from data.seoul, users share their arrival station to other users who are currently on board or on the train. By sharing their arrival stations, users can know when and where vacant seats will occur, thereby enabling a more comfortable commuting life.

text generation, GPT, fine-tuning

🎉Won 3rd prize!

Me and my teammates participated in the writing contest Bookathon with pre-trained GPT. The task was to fine-tune the pre-trained GPT provided by Minds Lab to generate an essay. About 5,000 essays in the 'travel' category were crawled to collect data at Brunch, a platform that posts various categories of essays. As the preprocessing stage was considered the most important in training process, we filtered and refined all the crawled data by ourselves.

Our key idea to improve the performance was sequentially feeding the data through two stages with learning rate decay. The ideal objective is as follows; At first, we want our model to learn essay styles with a large amount of data. The second was to train them to produce high-quality writings with highly preprocessed data. In addition, employing the learning rate decay at feeding the second data leads to a great loss reduction, enabling the model to not fall to the local minimum but to the global minimum. Finally, about 26 pages of essay was generated!

Data analysis, SQL

📖In the database management2 class in 2020 Fall semester!

In the early days of COVID-19, patients' infection information was indiscriminately scattered, causing confusion among users. This application provides five functions that allow users to access the information about COVID-19 in South Korea; mass infection cases, current COVID-related social distance policies, regional confirmed cases, risk level by each region, and up-to-date information updates.