cv

Importance Sampling Based Federated Unsupervised Representation Learning

IEEE

Published work on a novel FL algorithm for unsupervised and personalized representation learning.

Software Engineer

Microsoft

Building impactful APIs for Microsoft Graph Teams, specializing in meetings, teams, and channels. Developed a CI/CD pipeline for a Business Intelligence data pipeline.

• Specialization in API development and DevOps
• Built APIs for Microsoft Graph Teams
• Developed a CI/CD pipeline for Business Intelligence
• On-call for large-scale systems, reliability, and optimization

Software Engineer Intern

Microsoft

Worked on ECS flighting system for Audio, Signal and MediaTA under OneDrive SharePoint.

• Implemented ECS flighting system for Signals
• Developed private preview tasks for specific tenants

Deep Learning Android Game Developer

Atletik Games

Developed Android games based on deep learning for Physical Education during Covid-19 lockdown.

• Integrated state-of-the-art pose and object detection models
• Deployed games to 3 schools with 500+ students

B.Tech. Thesis

Federated Learning with unannotated data

• Developed a novel FL algorithm for unsupervised and personalized representation learning, later published in ICASSP 2024.
• Used Variational AutoEncoders to formulate clients' unlabelled and non-IID data to gaussian distributions.
• Aggregated distributions using mean-field formulation. Used importance sampling based weights for calculating reconstruction loss.
• Is-FedVAE outperformed SoTA FL baselines and demonstrated faster convergence and higher accuracy.

Independent Project

Plug-and-Play Platform for COVID-19 Resource Allocation

• Developed a reinforcement learning (RL) based platform for efficient resource allocation during the COVID-19 pandemic.
• Designed a generalized platform that could optimize the allocation of any resource by taking input datasets and considering relevant columns.
• Applied actor-critic methods to optimize allocation, ensuring a scalable, plug-and-play system for varied use cases.

Independent Project

Vehicular Planning using Deep Reinforcement Learning

• Worked on optimizing Autonomous Agent Vehicle communication for local and extended view under bandwidth constraints using Deep Reinforcement Learning methods.
• Surveyed literature and designed an architecture for multi-modal learning giving multiple outputs. Curated a reward system for efficient learning.

Research Project

SeekSuspect

• SeekSuspect is an interactive facial image retrieval model that uses visual memory as input to retrieve faces from a large-scale database.
• Fine-tuned pre-trained ResNet and VGG-16 models, improving accuracy from 82% to 86% by enhancing efficiency and embedding extraction.
• Evaluated and compared various clustering algorithms for embedding extraction, improving model performance and retrieval accuracy.

Bachelor

Indraprastha Institute of Information and Technology Delhi (IIITD)

Computer Science and Applied Mathematics

• Deep Learning
• Artificial Intelligence
• Reinforcement Learning
• Data Mining
• Algorithm Design and Analysis