Designed and developed data processing pipelines for multiple click-to-conversion ratio prediction ML models using a MapReduce based framework. Made the pipelines scalable, extensible and fault-tolerant for multiple models and gained significant performance gain. Currently working on integrating an end-to-end pipeline management system for automating various pipeline tasks.

Impact: Previous approaches were not fault-tolerant and scalable and caused training workflow errors as the data processing pipeline is upstream for model training. Minimized the rate of training workflow failures by almost 30% and gained 50% performance. improvement in training data generation process.

Engineer in Customer Experience team of Amazon Fresh (world wide grocery delivery service), goal is to build software to recommend grocery items to users on Amazon web pages. Designed, developed and released a scalable and extensible precompute system from scratch using AWS infrastructure to compute product data required for multiple recommendation projects across Amazon Fresh.

Developed a recommendation widget titled “Savings Maximizer” providing product recommendation with lowest price per unit, customer brand preference at millisecond latency, enabling customers to maximise their savings in grocery purchase.

Impact: Extensibility of precompute system helped in driving Supersaver, Inline upsell, Savings maximizer recommendation projects to success, saving an effort of approximately 3 months for each project. Savings Maximizer widget helps the customers to maximize their savings by recommending items with lowest price per unit and is estimated to create significant increase in order product sales in terms of revenue for Amazon Fresh.

Engineer in the AcrobatJS team at Adobe, goal is to build a fully functional Adobe Acrobat on the web. Worked in the field of mobile optimization, developed zooming capability in slide mode PDF view, onboarded inline search and page scrubber on Acrobat web.
Technologies: React, Javascript, HTML, CSS

Machine learning intern in Adobe Exchange team (exchange.adobe.com). Built a recommendation system to recommend product extensions to customers using a siamese positive-negative pair neural network for Adobe Exchange Platform. Achieved an accuracy of 82 percent in generating recommendations to users. Built a generic model adaptable to different platforms of Adobe. (PPO offered)
Technologies: Python, Tensorflow, Java

Working under the guidance of Prof. Yoav Freund and Prof. David Kleinfeld in Neurophysics department to extend Google's version of neuroglancer to support polygon and volume annotations, currently only line, point annotations are supported. Neuroglancer is a WebGL-based viewer for displaying arbitrary (non axis-aligned) cross-sectional views of volumetric data. With the addition of polygon annotations, researchers will be able to better identify and represent the cross-sectional areas of interest in the volumetric data and the annotations can in turn be fed to train ML models to automate the process of annotating volumetric data. Github: link Technologies: Typescript, Javascript, HTML, CSS

Implemented a real-time comment system similar to google docs comments for GraphSpace platform using concepts of websockets/network programming enabling users to add comments as a means to discuss or share ideas with their collaborators for a given graph.
Blog summarizing my work.
Technologies: Python, Javascript, HTML, CSS

Full time honours student at Computer vision lab under the director of IIIT (Prof P.J.Narayanan). Developed an unsupervised algorithm to learn image representation in style space which can be used for image retrieval. Research paper was accepted at WACV 2020. Performed extensive research and developed a system using deep learning to perform view extrapolation on real world images
Teaching assistant for Optimization methods(Spring 20),Operating Systems(Fall 18)and Machine Learning(Fall 19) courses.
Technologies: Python, Javascript, HTML, CSS, Pytorch, Tensorflow