Full-Stack Software Engineer, May 2023 - Present

Collaborated with the Washington State Department of Licensing to develop an application assisting teen drivers in their path to mastering driving skills.

The application has been rolled out to over 15,000 teen drivers in the state of Washington.

Worked on the front-end (interactive training and testing modules), back-end (API design, database design, user authentication) and application deployment.

Software Engineering Intern, Heart Failure, Jan 2024 - May 2024
Data Science Intern, Global Data Science and Analytics, June 2023 - Dec 2023

Implemented a data analysis pipeline to identify proxy measurements for assessing the efficacy of Deep Brain Stimulation in Parkinson’s and Essential Tremor patients to reduce costs for post-market trials.

Developed interpretable and predictive machine learning models to forecast heart failure stages by leveraging both insurance claims and EHR data.

NLP Engineer, Nov 2022 - May 2024

Working on ASR systems for over 700 low-resource languages (to be integrated in a language learning app)

Building an open-source Universal Voice Activity Detection system trained on a variety of large-scale datasets.

Full-Stack Software Engineernig Intern, Equities Technology Division, May 2021 - July 2021

Built a Client Simulator, mimicing client-side trading applications to test Citi's Order Management System.

Implemented the FIX protocol stack for communication with external APIs and interfaces.

Built the JUnit Test Suite for various FIX message scenarios.

A novel graphical password algorithm that would be immune to brute force attacks and at the same time, easy to remember.
During registration, the user selects a sequence of images (and within each image, a sequence of pixels/cells) that would be encrypted and stored in the database. During log-in, the user would then select his/her own images from a sequence of random images. Once the correct sequence is entered, the user will be authenticated.
Additionally, we use image captioning with named entity recognition (used to mask the keywords) to provide hints to the user in case the password was forgotten.
This project was part of the Smart India Hackathon (SIH) 2022.

A Semantic Search Engine for research papers on COVID-19 in various categories such as symptoms, influential factors, similar diseases and viruses, etc. This is a solution to the COVID-19 Open Research Dataset Challenge (CORD-19) on Kaggle.
The BioBERT model is used to generate sentence embeddings to encode semantics into the query and data.
Faiss, an open source library for efficient similarity search and clustering of dense vectors maintained by Facebook Research is used to efficiently compute the similarity scores of thousands of documents with respect to the query.

There exist a large number of meetings in the form of university lectures, business conferences, academic talks, interviews, etc. Each contain enormous amount of text in the form of conversations between multiple participants with varying levels of importance or influence. It is important to not only summarize the transcripts but also take into account which participant says what or what role they are assigned).
The proposed model builds on the HMNet model by Microsoft Research. We use a transformer based architecture along with a clustering module to rank the top 'n' sentences and make the summary attend to it in the decoder.
A simple attentive gating module is also introduced to suppress unimportant turn information.
Universal sentences embeddings are used to represent the text and role vectors to represent each speaker uniquely.

A platform, hosted using Amazon Web Services (AWS) to read and process any type of document, accurately extracting text, handwriting, tables and other data without any manual effort.
For this, Textract was used which is an Optical Character Recognition (OCR) service offered by Amazon. It can extract structured and unstructured information from images, PDFs, etc.
Amazon’s S3 bucket along with DynamoDB was used to store all the forms and associated data.

This is a solution to the IEEE PHM 2012 Prognostic Challenge. It focuses on the estimation of the Remaining Useful Life (RUL) of ball bearings, a critical problem among industrial machines, strongly affecting availability, security and cost effectiveness of mechanical systems.
We use a two-stage data-driven process for predicting the RUL using a single statistic derived from 14 time-domain features of the observed vibration signals. The degradation process is modeled as a Wiener process and the prior parameters of the underlying linear state-space model of degradation are continuously updated.

An online platform for students, teachers, professionals and enthusiasts to share educational resources, articles, ask and solve others' queries, so that they can learn and develop skills together. Users can chat among each other and create subforums and communities for a more collaborative discussion.

Implementation of a 1D Convolutional Neural Network for Sentence Classification based on the paper by Yoon Kim[2014].