• Led a team of 3 to design and develop a sentiment analysis system for classifying 53,000+ mental health-related text entries into predefined categories using deep learning.
• Contributed extensively to model development and training, implementing and comparing three advanced architectures Multi-Channel CNN, LSTM with L2 Regularization, and BERT with XGBoost — achieving a best test accuracy of 93.74% using Multi-channel CNN.
• Applied NLP preprocessing techniques including stopword removal, stemming, tokenization, Word2Vec and BERT embeddings to optimize model input and improve classification accuracy.
• Conducted 5-fold cross-validation to ensure model generalizability, with CNN and LSTM models achieving average accuracies of 93.58% and 92.90%, respectively.
• Performed hypothesis testing (paired t-test) to statistically compare CNN and LSTM performance, concluding no significant difference (p = 0.0733), despite CNN’s higher accuracy.
• Developed an interactive tool using ipywidgets that allows users to input text and receive real-time predictions from the selected model (CNN, LSTM, or BERT).

Project Link

• This project involves the image classification of traffic signs using a Convolutional Neural Network (CNN) trained on the German Traffic Sign Recognition Benchmark (GTSRB) dataset
• The dataset contains images in .ppm format (Portable Pixmap), which are resized to 30x30 pixels for uniform input dimensions.
• Achieved 98.70% test accuracy and 97.65% training accuracy in classifying 43 traffic sign categories using a custom Convolutional Neural Network (CNN).
• Enhanced model performance by incorporating Batch Normalization after convolutional layers and tuning dense layer architecture for better feature representation and generalization.
• Improved training stability and convergence by reducing the learning rate of the Adam optimizer to 0.0001 and eliminating redundant dense layers.
• Resolved overfitting issues through dropout regularization, and normalization techniques, increasing both model accuracy and reliability.

Project Link

• Designed and implemented a recommendation system that suggests movies based on user input, leveraging a dataset with over 60,000 records.
• Applied supervised learning techniques to train the recommendation model, improving the accuracy and relevance of recommendations.
• Built an interactive user interface using IPYWidgets, allowing users to search for movies and receive real-time recommendations in a user-friendly manner.
• Conducted extensive data preprocessing and cleaning using Pandas to ensure high-quality input for the recommendation algorithm, handling missing values and inconsistencies.
• Calculated the TF_IDF Score for each documenet term and also determined the cosine-simililarity score to identify the similar ones

Project Link

• Created a robust web scraping pipeline to collect comprehensive Premier League statistics, automating data extraction from multiple sources.
• Utilized Beautiful Soup and Requests to scrape data from websites, ensuring a diverse and rich dataset for analysis.
• Incorporated time delays and rate limiting to avoid overloading target websites and to comply with web scraping best practices.
• Developed error handling mechanisms to manage potential issues during scraping, such as connection errors and missing data, ensuring the robustness of the data extraction process.
• Saved the processed data into a CSV file with over 4,000 rows, implementing efficient data storage practices for ease of access and further analysis.

Project Link

• Designed and implemented machine learning models to classify NBA players into positions (SG, PG, SF, PF, C) based on their performance statistics.
• Utilized Pandas and Numpy to preprocess and clean the dataset, handling missing values, normalizing features, and preparing data for model training.
• Implemented and compared several classification algorithms, including Support Vector Machine (SVM), Naive Bayes, and Decision Tree, to determine the best-performing model.
• Achieved a significant improvement in model accuracy, with a training set score of 63% and a validation test score of 65% using SVM, by optimizing the feature set.
• Employed 10-fold stratified cross-validation to assess the robustness and generalizability of the models, resulting in an average score of 64%.

Project Link

• Designed and built a cross-platform e-commerce application for item exchange using React Native, ensuring a seamless experience on both iOS and Android devices.
• Designed and implemented a user-friendly interface using React Native components, HTML, CSS, and Tailwind CSS, ensuring a responsive and visually appealing experience across devices.
• Leveraged Google Firebase’s real-time database capabilities to enable real-time updates and synchronization of item listings, improving user engagement and interaction.
• Followed Agile practices for project management and development, including iterative development, sprint planning, and regular stand-ups, to ensure timely delivery and adaptability.
• Used Jira for tracking project tasks, managing user stories, and monitoring progress, leading to improved organization and project management.

Project Link