My Projects
A selection of my work in Machine Learning and Data Science.
Object Detection with DETR on KITTI Dataset
January 2024 - Present
Developed and fine-tuned a DEtection TRansformer (DETR) model for object detection tasks using the challenging KITTI dataset. This involved creating a custom PyTorch dataset class for KITTI and adapting the model's architecture for optimal performance.
- Implemented the KITTI API by transforming labels to a COCO-like format.
- Customized the final classification layer of a pre-trained DETR model to align with KITTI's label mapping.
- Achieved a 12% increase in detection performance (average AP) through extensive hyperparameter optimization.
- Enabled seamless model evaluation with enhanced performance metrics.
AI-Powered Coupon Discovery Chatbot
Personal Project
Dual-mode web chatbot that lets users search 2,510+ brands and categories from SimplyCodes. Includes a fast Basic Search API and an OpenAI-powered AI Chat mode with semantic search and conversational memory.
- FastAPI backend with Basic and AI endpoints; modern web UI with mode toggle.
- Playwright scraper discovers categories and scrapes coupons, building a 3-level tree with retry logic and anti-detection (19 main categories, 1,428 level-2, 1,082 level-3).
- LangChain-powered semantic retrieval using embedding-based vector search with conversational memory and retrieval-augmented generation for smarter recommendations.
Distributed Databases for Amazon Products
USC Viterbi School of Engineering Project (Group Project)
This project harnessed MongoDB's capabilities to architect a distributed database system tailored for Amazon products. It focused on ensuring optimal performance, scalability, and enhanced user experience for e-commerce data management through data distribution strategies and intuitive interfaces.
- Designed a data schema for essential product attributes and utilized 'asin' for sharding across MongoDB nodes using hashing techniques.
- Implemented data cleaning and loading from an Amazon Fashion JSON dataset, merging with metadata.
- Developed a database manager interface using Tkinter, providing comprehensive CRUD operations.
- Created a user-facing web interface with Streamlit for intuitive product querying, filtering, and navigation.
- Established a dedicated database node for backup and recovery processes to enhance data durability.
Deep Learning for Commodity Price Prediction
University of Liverpool - COMP390 (2022/23)
This project focuses on the prediction of commodity prices using deep learning models, offering a comprehensive examination and performance analysis of various models on this task. The primary goal was to implement and evaluate the effectiveness of several machine learning and deep learning models in predicting the close prices of commodities based on historical price data.
- Implemented and compared Linear Regression, Support Vector Regression (SVR), Long Short-Term Memory (LSTM), and CNN+LSTM combined models.
- Analyzed the impact of additional features (e.g., moving averages, RSI) on LSTM model performance.
- Explored feature extraction using CNNs for different timeframes to feed into an LSTM for prediction.
- The study found LSTM to have relatively the best performance among the implemented models for this specific task and dataset.