Projects



TransRecG: Transformer Based Sequential Recommendation System with Graph Embeddings (2023)

  • Sequential Transformer-based recommender system augmented using a Knowledge Graph to improve the recommendations.
  • This model captures both the higher order connectivity information, and the sequential order with which the users interact with the movies and this improves the performance.
  • This model uses a RGCN-based NGCF model trained on a user-movie-attribute Knowledge Graph to capture the higher order connectivity features and a Behavioral Sequence Transformer to capture the user's sequential information.
  • Received the Best Project in the Class award.
BERTG: Graph Augmented Transformer for Text Classification (2023)

  • Combines the local structural relationship between input documents learnt by the Graph Neural Networks with the contextual embeddings learnt by transformer-based models.
  • BERTG surpasses the performance of both BERT and GCN when used individually. And it achieves better results compared to BERT and GCN for smaller datasets as it requires lesser training data.
Emotion Classification from Text (2023)

  • Adviced by Dr.Bruce Walker
  • Building an AI-powered mental-health monitoring system using which a person can perform daily check-ins.
  • I worked in the Voice team and used NLP and Audio analytics techniques to analyze the user’s response from the daily check-ins. For text-based emotion analysis, I am using Transformer models like T5 etc.
  • This project is being developed in partnership with Primaface
Multimodal Sarcasm Detection (2022 - CS7641 Coursework)

  • Predicting if a video is sarcastic or not based on the video (visual content), voice and text (transcriptions from the video)
  • Built and evaluated the performance of several ML algorithms like Naive Bayes, SVMs, DNNs, Gaussian Mixture Model etc. The model were built for individual modalities and also a combination of two or all the modalities.
Data Visualization - Global Warming Dashboard (2022 - CS6730 Coursework)

  • Designed a dashboard using Tableau to visualize the effects of Global Warming.
VisBuddy - Smart Assistant for the Visually Challenged (2021)

  • A smart assistant for the visually challenged built using Tensorflow 2.4 and deployed on Raspberry Pi 4B.
  • It does 4 information Processing tasks namely: Scene Recognition, Object Discovery, Reading Assistant and Newsreader.
  • The Scene Recognition module uses Attention based Image Captioning algorithm proposed in the "Show Attend and Tell" paper with some minor modifications.
  • The Object Discovery Module uses RetinaNet and certain novel heuristics to determine the location of an object and guide the user to that.
  • The Reading Assistant uses Google Vision API and certain novel text processing algorithms to read the text in the user's view.
  • The Newsreader uses web scraping for getting the news articles and reads it to the user.

Paper Pro (2019)

  • A Django website for conference management and journal publication.
  • It has features to upload, review and update the submissions.