I am an MS in CS student at Georgia Institute of Technology, Atlanta campus. I am currently pursuing my specialization in Machine Learning and have taken the following courses as part of my specialization:
Working with the perception team to build a curated dataset for scene text detection and recognition in a warehouse setting, and implementing an efficient state-of-the-art algortihm for finetuning on the same.
Worked with the Google Assistant team to build an efficient debugging tool using C++ for the query understanding pipeline that ascertains machine-understandable intents for a given user speech query.
I worked as teaching assistant for the course CS 6515, Advanced Algorithms in Spring 2023. Further, I took up this role for the Machine Learning course (CS4641/7641) in Fall 2022. My job responsibilities included taking office hours, answering doubts on student-forums & classes, and grading assignments/exams/projects.
I worked as a backend developer for the e-commerce channel of Walmart, South Africa and Walmart, India . I contributed to various crucial features such as stock management, payments and quotes on the website leveraging web technologies- spring MVC (HYBRIS), HTML, CSS, JSP. As part of the Walmart India Team, I worked on building a recommender system that suggested a variant product to customers according to their location. I used Pytorch to build the machine-learning model, React Native to create a User Interface, "Element.ai," an E2E platform by Walmart Labs for model training and deployment. I was also a part of a crowdsourcing initiative to find cost-efficient solutions to track inventory at Walmart Stores using Object Detection.
I worked with the vision and language lab on the Visual Question Answering and Visual Dialog Task. I leveraged PyTorch to implement various VQA frameworks such as Stacked Attention Network (SAN) & VQA Counting, and studied their attention module. Conclusively, I, along with my team, successfully devised a generalisable algorithm to supervise and improve Attention module in any deep learning task. We got out our findings published in IEEE WACV’21.
I worked with 3 other summer interns to integrate the Walmart Canada website with the Google Home device. I used Google Dialog Flow to process user speech and recognize the user’s intent and employed Node JS Webhook to connect the Dialog Flow App with Walmart Canada APIs.
I, along with my team members, compared the performance of DL architectures such as MultiModal Bitranformers and Concat-Bert for hateful meme detection. We qualitatively studied the shortcomings of these architectures by analyzing GRADCAM heatmaps of false-negative and false-positive samples.
We used PCA to obtain most salient image features and employed ensembling techniques to combine the advantages of various DL models, achieving a significant improvement in the accuracy of baseline architectures.
[Code] [Website]
I, along with my team members, compared the performance of variations of Se2Seq models on Hindi Text Summarization. We further examine the performance gain with beam search with controlled patience and also examine improvements in the model with Fast Text Embeddings. We qualitatively studied the the architectures by performing attention visualization.
We also finetune Multilingual T5 on Hindi Text Summarization Dataset.
[Code] [Report]
COVID-19 infected more than 10,000 people in South Korea. In this project, We employed DS4C South Korea Patient dataset (announced by Korea Centers for Disease Control & Prevention) to build the contact trace using SoTA algorithms and determine likely infected people which were not reported.
[Code]
[Report]
In this project, we proposed an encoder-decoder model to create a better sentence level embedding and evaluated the model on the task of Paraphrase Question generation. The semantics were captured by a pairwise decoder that enforces encodings of similar sentences to be close to each other. Further, the syntactics of sentences were captured by stacking a GCN over LSTM states.
[Code]
