VU HAI DANG

Technologies:

• PyTorch: Built and deployed the DnCNN model using PyTorch, leveraging its tensor operations and GPU support for efficient image denoising.
• Flask: Developed a web application with Flask to create an interactive interface for uploading noisy images and displaying denoised results.
• Pillow (PIL): Utilized Pillow for image processing, handling loading, format conversion (RGB, grayscale), and saving of images.
• Torchvision: Employed Torchvision's transforms module for seamless image-to-tensor conversion and back, integrating with the DnCNN model.

Technologies:

• LangChain: Utilized LangChain for building the Retrieval-Augmented Generation (RAG) pipeline, enabling efficient document retrieval and question-answering capabilities through integration with vector databases and language models.
• Transformers (Hugging Face): Leveraged the Hugging Face Transformers library to load and fine-tune the Llama-2-7b-chat model, and to manage tokenization and text generation for precise and context-aware responses.
• ChromaDB: Employed ChromaDB as the vector database to store and query domain-specific and Q&A data, facilitating fast and accurate similarity searches for relevant document retrieval.
• Streamlit: Used Streamlit to develop an interactive web-based user interface, providing features like conversation history management, model configuration, and real-time chat interactions.
• PyTorch: Integrated PyTorch for GPU memory management and optimization, ensuring efficient handling of large language models during inference and fine-tuning processes.
• HuggingFace Embeddings: Implemented sentence-transformers/all-MiniLM-L6-v2 from the HuggingFace Embeddings library for text vectorization, enabling robust semantic search within the vector database

• Car detection: Trained a YOLOv8 model to detect vehicles in traffic images.
• Image stitching: Created a panoramic image by stitching multiple images using OpenCV.
• Font classification: Developed a solution to classify 5 font groups.
• Flood classification: Trained an AI model to predict flood levels based on text, achieving a score of 0.92 and ranking in the top 10 in the Deep Learning (DPL) competition on Kaggle.