Orchestrated a legal document RAG system leveraging the Pinecone vector database for semantic search, FastAPI for scalable backend services, and integrated LangChain with Llama 3.3, achieving 75% more accuracy in document retrieval and contextual response generation.

Developed a Conditional GAN to generate synthetic images for underrepresented classes, improving dataset balance and model performance with comprehensive data handling scripts.

Developed a U-Net architecture for image segmentation, achieving a validation accuracy of 98% and an Intersection over Union (IoU) score of 95%.

Implemented the vanilla VQA system using PyTorch, replicating the research paper VQA: Visual Question Answering with data handling scripts for downloading, preprocessing, and preparing image and question datasets and for training the model.

Compared the performance of QNNs and CNNs, achieving 80% test accuracy for CNN and 65% test accuracy for QNN. Highlighted computational trade-offs, with QNNs taking ~20 seconds per epoch compared to CNNs' <2 seconds.