My Development Journey

Used the Speech T5 Model as the base, fine-tuning it with the Lithuanian language (lt) from the Facebook/VoxPopuli dataset. This fine-tuning process enhances the TTS model's performance by adapting it to specific domains, accents, or speaker voices.

Developed a web application to extract English and Hindi text from uploaded images using OCR and display results in plain text or JSON format. Integrated a document search feature for keyword detection, highlighting matches, and providing contextual feedback.

Developed a machine learning model using H-Bagging, an enhanced ensemble classification approach, to accurately predict Parkinson's disease from voice data. The model utilized heterogeneous classifiers (SVM, Naive Bayes, KNN, Decision Trees, and Logistic Regression) and achieved superior accuracy by focusing on misclassified samples through a novel M-Bootstrapping technique

An AI-powered image analysis web app that provides both text and audio output, with additional features like output translation, custom image analysis based on user input, and image generation. Integrated a zero-shot learning object detection model using the Gemini API, capable of identifying objects beyond its training set, and leveraged Retrieval-Augmented Generation (RAG) to enhance detection accuracy by incorporating external knowledge. Utilized few-shot encoding to customize the response structure, ensuring adaptability to diverse categories.

Predicted the outcome of football matches for top leagues based on the last 25 years of results using a comprehensive preprocessing pipeline that included both numerical and categorical feature transformations. Trained separate RandomForestRegressor models for predicting home and away team goals, achieving an RMSE of 1.03. The model incorporated various features such as team form and historical match data to enhance prediction accuracy. Additionally, implemented cross-validation and hyperparameter tuning to optimize model performance and ensure robustness.

Developed an AI-powered trip planner offering unbiased travel recommendations and a virtual tour feature. Integrated an Augmented and Modified AI model to provide personalized travel plans, allowing users to customize their holidays independently of travel agencies. The system analyzes user preferences, budget constraints, and travel history to generate tailored itineraries. Implemented 2-Factor Authentication using Backend as a Service (BaaS) for enhanced security, ensuring user data protection and secure access. Leveraged serverless backends to ensure scalability and protection against data breaches and server attacks, providing a robust and reliable platform for users to plan their trips with confidence.