Hello, I'm

My Projects

• Dynamic Personal Website: Next.js, React, Tailwind CSS

  This dynamic personal portfolio was built with Next.js, React, and Tailwind CSS, featuring a sleek, responsive design and smooth animations via Framer Motion. The project showcases modern front-end development practices, state management, and a focus on UI/UX excellence.

• Stock Price Trend Prediction (VAE + CV)

  Leveraged Variational Encoder-Decoder (VED) and candlestick chart images to predict 5-day stock price trends based on the past 20-day SSE50 index. Combined deep learning with financial visualization to explore an image-based approach to forecasting—an innovation over traditional raw data models.

• Lightweight ResNet for CIFAR-10 with 2.7M Params

  Designed a compact ResNet variant with only 2.7M parameters, achieving 94.06% accuracy on CIFAR-10. Outperformed standard ResNet-18 (11.2M) by balancing depth and efficiency—an effective solution for small dataset generalization.

• Parameter-Efficient Text Classification with LoRA

  Fine-tuned a RoBERTa model using Low-Rank Adaptation (LoRA) on the AGNEWS dataset with fewer than 1M trainable parameters. Achieved 87.4% accuracy, demonstrating that LoRA can deliver competitive performance with minimal computational cost in real-world NLP tasks.

• Adversarial Robustness of ResNet-34 on ImageNet

  Studied ResNet-34’s vulnerability to FGSM, PGD, and patch attacks. Demonstrated drastic accuracy drops and evaluated the transferability of adversarial examples to DenseNet-121 and ViT-B/16, highlighting security weaknesses in deep image classifiers.

• Tremor & Dyskinesia Detection on STM32

  Built an embedded system on STM32 to detect Parkinsonian tremors (3–5Hz) and dyskinesia (5–7Hz) using accelerometer/gyroscope FFT analysis. Implemented LED indicators to signal movement type and intensity, enabling medication-level tuning without terminal output.

• JFK Airport Lane Traffic Detection and Analysis

  A computer vision project focused on image recognition and object detection using yolo-v8 computer vision models.

• Evaluating the Impact of Tariff-Sentiment on Excess Returns

  Built a fully automated Python event-study pipeline integrating CRSP, Fama-French, and FinBERT data to quantify the market relevance of tariff sentiment in corporate earnings disclosures.

• Moodify: Emotion-Aware Music Recommendation System

  Implemented a scalable emotion-aware music recommendation pipeline using PySpark, HuggingFace Transformers, and Flask. Processed the Spotify Million Playlist Dataset to extract popular tracks, built a fault-tolerant lyrics scraping and cleaning system, and applied DistilRoBERTa-based emotion inference. Translated emotion predictions into mood-based playlists and exposed results via a responsive web application.