I'm an M.S. student at the POSTECH Graduate School of AI, researching in the ML Lab. My interests center on LLM agents, machine learning for NLP, and generative AI.
An M.S. student at POSTECH, working on LLM agents and machine learning for NLP.
I'm an M.S. student at the POSTECH Graduate School of AI, in the Machine Learning Lab, where I am supervised by Prof. Dongwoo Kim. My research interests are LLM agents, machine learning for NLP, and generative AI.
Before POSTECH I studied Computer Science at Ewha Womans University and interned on the AI team at Lotte Homeshopping, where I built embedding-based product-matching engines over hundreds of thousands of products and a conversational data-analysis tool. That applied-ML background is the foundation I now build research on.
Built KR-SBERT embedding + DBSCAN clustering pipelines to identify identical and similar products across internal and competitor catalogs, improving the accuracy of sales forecasting. Similarity combined product name, category, brand and price with tuned weights.
Automated the reading of QA documents with LLMs on AWS Bedrock. Designed a chunking strategy that fixed table-reading failures in a RAG QA chatbot, and built a Selenium / PyMuPDF pipeline that extracts reference numbers from PDFs, verifies test-report authenticity online, and assembles evidence screenshots — shipped as Lambda APIs.
Prototyped a Dockerized BI assistant that unifies natural-language SQL queries with general LLM chat. A routing flow sends each query to a Text-to-SQL path or a chat path, runs it against MySQL, and post-processes the result into a user-friendly answer.
A hybrid recommender combining content-based filtering over perfume notes and descriptions with collaborative filtering over user preferences. S-BERT embeddings indexed in FAISS for fast similarity search; user "likes" stored as a Neo4j graph for similar-user recommendations. Deployed with Streamlit.
A platform that builds children's logical reasoning through quizzes. GPT-4o mini generates and grades quizzes; a rubric-based chain-of-thought prompt returns per-criterion scores with reasons. A Korean-specialized Llama model was fine-tuned with LoRA + 8-bit quantization for efficient training and memory use.
A personalized scholarship-recommendation platform centered on AI and backend development. Generative AI drives the recommendations, and an optimized filtering algorithm cut processing time by ~97.3%. Built on Django, MySQL and Redis, containerized and deployed on AWS.
A CNN classifier that distinguishes benign from malignant skin lesions on the Kaggle ISIC 2024 dataset. Data augmentation addressed severe class imbalance, and a fine-tuned ResNet-50 lifted classification performance.
A deep-learning classifier for 20 Pokémon species, covering the full pipeline from data preprocessing through systematic comparison of several architectures. MobileNet was selected as the most efficient model and fine-tuned, reaching 90.41% test accuracy.
I'm always glad to talk about LLM agents, ML for NLP, or generative AI — research collaborations, internships, or just ideas. The fastest way to reach me is email.