Projects

[M.S. Thesis] AlphaSS : Protein Structure Prediction with Disulfide Bond Information

[M.S. Thesis] AlphaSS : Protein Structure Prediction with Disulfide Bond Information

AlphaFold2PyTorchDeep LearningProtein Structure PredictionBioinformatics

Enhance AlphaFold2 by integrating disulfide bond embeddings and loss to improve prediction accuracy, especially in low-MSA scenarios.

2023.03 - 2025.02

1. Objective

  • Enhance AlphaFold2 by integrating disulfide bond embeddings and disulfide loss to improve prediction accuracy, especially in low-MSA scenarios.

2. Approach

  • In-depth Analysis: Conducted protein structure data analysis and feature extraction to identify key factors influencing structure prediction.
  • Model Optimization: Developed and optimized a modified AlphaFold2 pipeline incorporating disulfide-specific features.
  • Benchmarking: Rigorously benchmarked performance improvements against standard datasets to validate the effectiveness of the proposed method.

3. Results

  • Significant Performance Gains: Observed a TM-score improvement of 2–3% under sufficient MSA conditions and 5–10% under insufficient MSA conditions.
  • Enhanced Recall: Disulfide bond prediction recall improved by 50–100% with sufficient MSA, and by 60–100% when MSA was limited.
  • Low-MSA Robustness: Validated that incorporating disulfide bond information particularly benefits prediction performance under data-scarce (low-MSA) conditions.
  • Conference Presentation: Findings were presented at BIOINFO 2024.
GitHubThesis
[LG Aimers 7th] Menu demand forecasting with N-HiTS & designed loss function

[LG Aimers 7th] Menu demand forecasting with N-HiTS & designed loss function

Time Series ForecastingN-HiTSLG Aimers 7thLLMHackathon

[LG Aimers 7th Hackathon finalist] Time series forecasting using N-HiTS and LLM

2025.08 - 2025.09

[LG Aimers 7th Hackathon finalist]

1. Log Transformation & Data Normalization

  • Log-transformed sales quantity targets for training stability (training/validation/prediction) → Inverse transformed for final results.
  • Applied MinMaxScaling (0-1) to other covariates per series (store x menu).

2. N-HiTS Based Modeling Strategy

  • Simple Temporal Split: Maintained time order.
  • Multi-Window Validation: Set a sufficiently long validation span (42~56 days) to test multiple 28-day input → 7-day forecast windows for average performance.
  • Consistency: Configured validation identically to the actual submission task ("predict next 7 days using last 28 days") to accurately assess overfitting/underfitting.
  • Covariate Alignment: Split covariates at the same time steps as targets to prevent mismatch errors.

3. Evaluation Metrics & Loss Functions

  • Adapted loss function approach due to different metrics in preliminary vs. final rounds.
  • 3.1. Preliminary Round: Weighted SMAPE by store. Used oversampling on specific high-weight stores (e.g., 'Mirasia') to learn weights.
  • 3.2. Final Round: SMAPE, NMAE, NRMSE, R-squared. Equal weights per store (no oversampling). Masked segments where target is 0 or negative as they are excluded from evaluation.

4. Cluster-Based Modeling

  • Identified menus with residuals > 10 in validation data.
  • Performed Spearman correlation-based cluster analysis on menus from the high-weight store ('Mirasia') with high residuals.
  • Classified into 3 clusters based on similar trend patterns.
DACONRelated Link (LG AI Research)Related Link (LG Aimers)
AI Parmacist Chatbot Service

AI Parmacist Chatbot Service

RAGOCRLangGraphLLMHackathon

[Upstage Global AI Hackathon finalist] RAG와 Langchain, Solar LLM Finetuning을 이용한 영양제 추천 서비스 개발

2024.08 - 2024.09

Finalist in Global AI Week Hackathon (Upstage AI Hackathon).

  • Objective: Develop an AI-driven chatbot to assist with supplement and medication intake.
  • Approach:
    • Designed and implemented chatbot functionality using LLMs, Langchain, RAG, and Gradio.
    • Optimized responses through retrieval-augmented generation (RAG).
GitHub
AI-driven Game Scenario Generator

AI-driven Game Scenario Generator

Prompt EngineeringDALL-ELLMHackathon

[스마일게이트 퓨쳐랩 AI 서비스 위클리톤 대상 수상] LLM Finetuning 및 Prompt engineering을 통한 게임 시나리오 제너레이터 개발

2024.07

1st place winner at Smilegate FutureLab AI Service Weeklython.

  • Objective: Create an AI-based game scenario generator for dynamic content creation.
  • Approach:
    • Applied GPT fine-tuning, prompt engineering, and DALL·E.
    • Developed an interactive UI using Gradio.
GitHubBlog PostRelated Link

Development of an Automatic Defect Sorting System Module for Paper Containers Using AI-based Vision Inspection

AIVisionYolov5Jetson Nano

NVIDIA JETSON NANO 및 Yolov5를 활용한 종이용기 불량품 자동 선별 시스템 모듈 개발

2022.03 - 2022.12
  • Developed an object detection model using YOLOv5 and built a classification system using Jetson Nano.
    • Constructed a dataset for detecting foreign substances in a local environment and fine-tuned a YOLOv5 baseline model.
  • Optimized the model for deployment on Jetson Nano by applying model compression techniques.
  • Achieved 95% accuracy in detecting foreign substances.
  • Presented a poster at KSME2022 (The Korean Society of Mechanical Engineers).

Tech Stack

  • YOLOv5
  • Jetson Nano
  • Python
  • CUDA
  • OpenCV
  • PyTorch
GitHubBlog Post