Researching Computer Vision under the supervision of Prof. Seunghoon Hong.

Organized and led a data labeling team to build a high-quality annotated dataset for training the cooking-robot segmentation model.

Developed a segmentation model for cooking robots to accurately identify safety-critical objects (pots, ingredients, oil, and human operators) in real-time.

Designed a custom evaluation metric to prevent real-world hazards by penalizing consecutive misclassifications of high-risk classes.

Optimized the model based on the custom metric, achieving ≥95% safety-compliant frames and limiting maximum consecutive errors to under 10 frames (<2 seconds).

Department Representative of ISE (과대표)

VOK (KAIST Broadcasting System), Executive Director of General Affairs (총무)

VOK (KAIST Broadcasting System), Announcer

Freshman Representative Speaker, KAIST Entrance Ceremony

President, Student Council

Graduated on the Domestic Track

Python, C, C++, Java

PyTorch, TensorFlow, OpenCV, CUDA, NumPy, Pandas, Seaborn

Jetson Nano, Arduino,  ESP32

Korean (Native), English (Fluent, iBT TOEFL 108)

Developed a machine learning pipeline using Finite Element Method (FEM) simulation and design specification data to predict tire manufacturing defects and optimize test production strategies.

Engineered custom features, including Target-Aware Correlation Drop and Mean-Difference Flags (MDFlag), and extracted structural statistics (e.g., pressure gradients) from complex 2D FEM data to isolate defect signals.

Optimized predictive stability on a highly imbalanced dataset (Defect ratio ~15%) by implementing a weighted ensemble model combining TabPFN (neural network for tabular data) and Random Forest.

Designed a profit-maximizing decision algorithm based on break-even analysis (Good: +100, Defect: -2000) to strictly penalize false negatives, strategically selecting the top 200 optimal tires to maximize Total Net Profit.

Addressed the high initial volatility and risk of standalone reinforcement learning models by implementing an ensemble technique that computes a weighted average of MVO and PPO asset distribution weights.

Conducted backtesting using 10 years of S&P 500 market data, achieving a cumulative return that outperformed the standalone Markowitz portfolio by 168% and the PPO-only portfolio by 336%.

Demonstrated superior risk-adjusted performance and robust risk management by maximizing both the Sharpe and Sortino ratios compared to traditional baseline models.

Engineered a hands-free, glasses-shaped wearable device designed to assist visually impaired individuals by conveying conversational emotions through a custom-coded vibration sensor.

Developed deep learning-based multimodal emotion recognition algorithms, achieving 98.55% accuracy in Facial Expression Recognition (FER) using the Xception model and 97.25% in Speech Emotion Recognition (SER) via Conv1D and LSTM.

Implemented an ESP32 microcontroller for real-time hardware activation via Bluetooth Low Energy (BLE) communication, controlled by a custom-built Flutter mobile application.

Designed an emotion integration logic that categorizes extracted emotions into positive, negative, and neutral states, cross-validating facial and audio inputs to ensure accurate and refined vibration feedback.