Sree Teja Nadella

In this project, we developed an Image Caption Generation system that uses artificial intelligence to generate captions for images. The system includes both frontend and backend components:

• Frontend Interface: Built a user-friendly interface using React.js and Tailwind CSS for uploading images and displaying AI-generated captions.
• Backend Processing: Used Flask to develop a backend API that processes uploaded images using a machine learning model built with TensorFlow or PyTorch to generate captions.
• User Authentication & Sessions: Implemented simple user authentication and session management using Flask-RESTful, Redis, and PostgreSQL (optional) for data storage.

Key features:

• Users can upload images from the frontend and receive captions generated by the AI model.
• Implemented responsive and modern UI with smooth animations using Framer Motion.
• Ensured efficient backend processing with Flask, TensorFlow/PyTorch, and Redis for caching.
• Leveraged PostgreSQL for storing user authentication data (optional).

This project highlights our ability to integrate machine learning with web development to build a fully functional application for real-time image captioning.

In this project, we developed Tomato: Online Food Ordering Application, a full-stack web platform that enables users to browse, order, and manage food items efficiently. The system includes both user and admin functionalities with a robust backend and a dynamic frontend:

• Frontend Interface: Built using React.js for a seamless user experience, allowing users to browse food items, view details, and add them to their cart.
• Admin Dashboard: Implemented an admin dashboard where authenticated users can manage food items by editing, adding, and deleting entries.
• Authentication & Security: Integrated user authentication and authorization using Node.js and MongoDB to ensure secure access to features.
• Database Management: Used MongoDB for storing food items and user data efficiently with a flexible schema.
• Backend Processing: Developed using Node.js and Express to handle user interactions, API requests, and business logic.

Key features:

• Users can browse food categories and add items to their cart through an interactive UI.
• Implemented a sliding mechanism for effortless navigation and an intuitive shopping experience.
• Admins can manage food inventory through a user-friendly dashboard with options to edit, upload, and delete items.
• Ensured responsive and modern UI with smooth animations and interactive components.
• Implemented secure login/logout functionality for both users and admins.

This project demonstrates our ability to integrate full-stack development with modern web technologies to build a user-centric and scalable online food ordering system.

In this project, we developed Disease Prediction from Patient Narratives, a system focused on predicting diseases using patient health narratives to facilitate early intervention and informed medical guidance. The system includes both traditional machine learning models and advanced transformer-based models:

• Dataset Collection and Preprocessing: Collected and preprocessed the Haodf dataset from Zhixiang et al., preparing it for model training and evaluation.
• Baseline Model Implementation: Implemented baseline models including TF-IDF + SVM, Logistic Regression, and RNN for initial disease prediction.
• Advanced Model Development: Developed and trained advanced models, including BERT variants, RoBERTa, DeBERTa, BioBERT, and MeDP, to enhance prediction accuracy.
• Performance Evaluation: Evaluated and compared the performance of baseline and advanced models to determine the most effective approaches.

Key features:

• Utilized the Haodf dataset, consisting of clinical text records with fields like wanted help, disease history, surgery history, allergy history, and medication usage.
• Implemented a preprocessing pipeline to clean and structure the text data, including removing URLs, HTML tags, special characters, lowercasing, and whitespace normalization.
• Fine-tuned pre-trained language models such as bert_base_uncased, bert_large_uncased, roberta-base, deberta-v3-base, and biobert-v1.1 using the Hugging Face Trainer API.
• Trained a lightweight transformer model, paraphrase-MiniLM-L6-v2, from scratch to evaluate the performance of smaller, more efficient models.
• Achieved enhanced performance with transformer-based models, outperforming baselines with a test accuracy of up to 0.8269 using BERT-Large-Uncased.

This project demonstrates our ability to apply NLP techniques, including transformer models, to real-world healthcare data for disease prediction, providing a foundation for future research and potential clinical applications.

In this project, we developed Retrieval-Augmented Generation (RAG) System with AWS Bedrock, a system designed to retrieve and generate responses using multiple PDFs stored as vector embeddings. The system leverages AWS Bedrock models for efficient and context-aware answer generation.

• Data Injection: Processed multiple PDFs by reading them from a specified folder, splitting them into chunks, and generating vector embeddings using Amazon Titan.
• Vector Storage: Stored the generated embeddings in a FAISS vector database for efficient retrieval.
• Querying Mechanism: Performed similarity search within the vector store based on the user's question to retrieve relevant document chunks.
• LLM Integration: Passed the retrieved document chunks along with a prompt to an LLM model, such as Claude or Llama 2, via AWS Bedrock.
• Response Generation: The LLM generated answers based on the provided context, ensuring accurate and relevant responses.

Key features:

• Utilized FAISS-CPU for efficient vector-based similarity search.
• Implemented a preprocessing pipeline to structure document chunks for optimal retrieval.
• Integrated AWS services, including boto3 and AWS CLI, for seamless cloud-based computation.
• Designed a Streamlit-based UI for an interactive user experience.
• Enabled flexible model selection, allowing responses from different AWS Bedrock-supported LLMs.

This project showcases the power of Retrieval-Augmented Generation, combining document retrieval and LLMs to generate informed responses, making it useful for knowledge-based applications.