Deviprasad Sunny

On my Local RAG Document Assistant project, I built a PDF question-answering application using document ingestion, HuggingFace embeddings, and Chroma vector search. It uses locally hosted Llama 3.2 through Ollama to generate document-grounded answers without paid LLM APIs.

I developed a LangGraph workflow for retrieval, relevance checking, query reformulation, and answer generation with source-page attribution, then exposed it through a FastAPI REST endpoint.

For my Real-Time Face Emotion Recognition project, I developed a Vision Transformer-based system using the FER2013 dataset and integrated the trained model with a live webcam pipeline using OpenCV. It achieved approximately 20 FPS inference during real-time prediction.

I also developed an NLP-based message spam classification system using Naive Bayes and LSTM, applying text preprocessing and TF-IDF vectorization. I evaluated it with accuracy, precision, recall, and F1-score, achieving an F1-score of 0.991.