KU-Chat is a domain-specific Q&A chatbot fine-tuned on KU web data to deliver accurate, context-aware answers for students and faculty. In this project, I designed and implemented a full data pipeline to build a KU-specific question-answering chatbot. Crawled university websites using Selenium and Airflow, extracting and cleaning over 1 million tokens. Generated structured Q&A pairs with OpenAI’s GPT-4o-mini and stored the processed data in AWS S3. Fine-tuned LLaMA 3.2 using PEFT-LoRA and 4-bit quantization, then tested performance with a Retrieval-Augmented Generation (RAG) setup.

This led to a 6% boost in ROUGE and BERTScore over the base model. The fine-tuned model delivered more specific, verbose, and context-aware answers, achieving over 90% accuracy on a custom KU evaluation set and significantly improving information access for students and faculty.

In this project, I developed a model to automatically detect and classify car damages. The model was deployed on an ESP32S board, enabling real-time car damage detection. Utilizing the CarDD dataset, I initially created a Convolutional Neural Network (CNN) model to perform detection and classification tasks. I then used the TensorFlow Object Detection API to train a MobileNetV2-SSD model for enhanced accuracy in identifying car damages.

This project enhanced my skills in model development, TensorFlow API, model optimization, embedded systems deployment, and managing end-to-end machine learning workflows.

The IMDb Reviews Analysis project uncovers patterns and trends in movie reviews, providing insights into how sentiments and ratings change over time. Through Sentiment Analysis, Temporal Analysis, and Correlational Analysis, we explore the dynamics of audience opinions and their evolution. Additionally, we developed a model to accurately predict the sentiment of a review.

I learned a lot about web scraping, text processing, and tokenization. I also gained experience with vectorization techniques and data visualization. Additionally, I developed skills in analyzing large datasets to uncover meaningful insights.

In this project, I developed an automated system using deep learning models to accurately detect and classify skin lesions. I trained custom CNN models and fine-tuned state-of-the-art models like ResNet & MobileNet to assist in early diagnosis and improve patient outcomes.

Through this project, I learned to implement custom datasets and dataloaders, build and train deep learning models, handle class imbalances, and apply data augmentation techniques, significantly enhancing my understanding of deep learning frameworks.

SoccerTact is a web application that provides comprehensive soccer analysis, including match based, team based and player based analysis. The application utilizes event data from StatsBomb's open data repository. We built a robust data pipeline to automatically extract data from the StatsBomb GitHub repository and store it in an SQL database.

I gained full-stack development experience, built a robust ETL pipeline, enhanced data analysis and visualization skills.

MAPKU is a web application developed to assist first-year students and staff in navigating the campus. Users can search for buildings by class number, mark multiple destinations for the fastest route, and view route details (estimated time, distance). The app also provides information about campus buildings.

This project enhanced my skills in web development, user experience design, and geolocation services integration. I also gained experience in optimizing algorithms for route calculation.

Developed a spam detection system using Bernoulli Naive Bayes from scratch in Python. The project involved data cleaning, feature extraction, and probabilistic modeling to classify SMS messages as spam or ham.

Gained skills in data preprocessing, binary feature matrix creation, and model implementation. Learned to calculate prior and likelihood probabilities, evaluate model performance using metrics like accuracy.

Built with Next.js, React, Tailwind CSS, and JavaScript, this site is my personal corner of the internet.

A fun side project that let me explore frontend development while showcasing my work — designed, developed, and deployed from scratch.