Projects
Implemented a document ingestion and preprocessing pipeline using Docling to convert PDFs into markdown and segmented large documents into manageable chunks with LangChain’s RecursiveCharacterTextSplitter for efficient embedding and querying. Integrated Mixedbread’s embedding model and open-source LLMs via Ollama, ensuring data privacy compliance by addressing security and ML requirements, and optimized Chroma vector store performance, enhancing answer relevance by 20% through custom prompt engineering. Engineered a robust testing framework with pytest, leveraging the same LLM used for querying to validate RAG system accuracy across 100+ predefined test cases, and designed a custom evaluation pipeline to rapidly detect response inconsistencies, ensuring continuous system reliability.
Developed convolutional neural network models using Pytorch and Keras to automatically predict locations based on landmarks in images, assisting in tagging photos that lack geolocation metadata. Engineered features from landmark images and optimized CNN hyperparameters, achieving 88% accuracy on a test set for classifying photos into different landmark location categories. Deployed the best-performing CNN model into a photo location prediction app, enabling the automatic tagging of images with relevant location information, demonstrating the practical application of the model.
Created a Python-based program designed to assist individuals with medical conditions such as Parkinson's disease, where typing can be challenging due to tremors or sudden, uncontrolled motor movements. By providing real-time spelling and grammar correction, it makes typing easier and more accurate. The program uses the Gemini language model from Google's Generative AI suite to intelligently correct text, simulating keystrokes to apply corrections.
Developed a Chrome extension that allows users to toggle the removal and restoration of YouTube Shorts, utilizing JavaScript, MutationObservers, and Chrome storage, achieving a 99% success rate in consistently removing Shorts. Optimized performance by reducing DOM access by 50% and implementing throttled MutationObservers, improving page load times by 15%. Actively used by 600+ Users Worldwide!
Headed a team of 5 in developing an interactive web application simulating critical algorithms such as Banker’s, Peterson’s, Dining-Philosopher’s, and Producer-Consumer, providing a practical learning platform. Integrated an explanatory chatbot to enhance algorithm comprehension for university juniors, and hosted the application on Netlify for accessible, hands-on learning.