Sep 2021 – Dec 2023 · 2 yrs 3 mos

• Threat detection for enterprise emails.
• Large Language Models (BERT, etc.) for text based advanced email attacks
• Auto Retraining Framework
• Multimodal model development

Sep 2021 – Dec 2023 · 2 yrs 3 mos

• Led launch of multimodal models reducing false negatives by 15–20%, resolving a critical FN surge and restoring metrics within OKRs in one month; enabled automated end-to-end retraining and deployment. Tech Lead for text-only email attack detection, delivering privacy- and compute-efficient NLP models; achieved 10% FNR reduction and an additional 8% via multimodal retraining with BERT embeddings. Scaled ML operations by adding hyperparameter tuning and multi-model support to the AutoRetraining Framework. Revamped Abuse Mailbox models, reducing manual review volume by 20% with no FN increase, and shipped targeted rule-based mitigations cutting customer-reported FNs by 10–15%.

Apr 2019 – Aug 2019 · 4 mos

• Open sourced pretraining and finetuning implementation for BERT language representation model using Azure Machine Learning Service. All the implementation details can be found in the Github repo (link below).

Jun 2017 – Sep 2021 · 4 yrs 3 mos

• Worked on GPT3 distillation for Github Copilot
• Bing LLM: Building end to end large scale NLP deep learning pipelines for Bing on Azure Machine Learning (Bing Core Relevance, Question Answering, etc.)
• Job Elasticity: Added support for Azure to PyTorch: https://github.com/pytorch/elastic/pull/55
• Experimenting with distributed training, performance benchmarking and optimization of large-scale ML models on AzureML.

Jun 2017 – Apr 2019 · 1 yr 10 mos

• Built smart add-in resiliency framework for Outlook Desktop

Sep 2016 – Apr 2017 · 7 mos · Stanford Center for Advanced Research through Online Learning

• Implementing a personalized, data-informed study guide for an individual learner anywhere in a MOOC using NLP. Raw material will be video attention profiles and learner success on the quizzes that immediately follow each video, forum posts, video closed caption files, and potentially outside material as well.

Jul 2016 – Sep 2016 · 2 mos · Santa Clara, CA, USA

• In collaboration with UC Berkeley Cloud Robotics and Automation group, advised by Prof. Ken Goldberg, Qiang Li and Nan Tian.
• Building a Cloud AI with deep learning, computer vision, and 3D visual SLAM to control a 7 DOF robotic arm for automatic and/or human guided grasping and decluttering of household objects by partially observed 3D object matching for grasping prediction using deep learning.
• Solving the problem of textureless image segmentation to recognize and track the chess pieces on a chess board for the robot to be able to play chess.
• Real-time 3D object tracking and pose estimation, image segmentation, de-cluttering, etc, using Xbox Kinect and Intel RealSense depth cameras
• Technologies used: OpenCV, OpenGL, Realsense SDK, Kinect, coding in C++

Jun 2016 – Sep 2016 · 3 mos

• Collaborated with UC Berkeley Cloud Robotics and Automation group. Implemented real-time 3D object tracking and pose estimation based on pointcloud segmentation using Xbox Kinect and Intel RealSense depth cameras. Visual feedback to control 7 DOF robotic arm for automatic grasping and texture-less decluttering of chess pieces.

Jan 2016 – Mar 2016 · 2 mos · Stanford University

• Advised by Prof. Fei Fei Li, we use machine learning and computer vision techniques to analyze Google street-view images of cars in hundreds of american cities, correlating them with relevant information such as household incomes etc. to be able to predict the next census. Currently focused on data analysis and visualizations on our car dataset, running baselines for fine-grained detection such as fast-RCNN and other aspects like domain adaptation.

Sep 2015 – Nov 2015 · 2 mos · Stanford University, CA

• The project is on the design and development of a distributed engine
• for interacting with IoT enabled devices (with a focus on domestic
• medical devices), and the associated domain specific language to
• program, with the goal of general public usability.
• The technologies we're using are primarily node.js, some db engine we
• haven't decided yet (mongo? some sql?), as well as the usual web
• technologies (html+css+js+jquery+bootstrap). The communication uses a
• custom JSON-based protocol on top of WebSockets. As our deployment
• platform, we target Android (and we use jxcore to run node.js modules)
• and private servers, including small ARM devices like the Raspberry Pi
• 2.

Jan 2016 – Jun 2016 · 5 mos · Stanford University

• Advised by Prof Amir Goldberg.
• Given that organizational culture is hard to observe, we develop a novel approach to assessing individuals’ cultural fit with their colleagues in an organization based on the language expressed in internal email communications. Drawing on a unique data set that includes a corpus of 10.25 million email messages exchanged over five years among 601 employees in a high-technology firm, we find that network constraint impedes, while cultural fit promotes, individual attainment. More importantly, we find evidence of a tradeoff between the two forms of embeddedness: cultural fit benefits individuals with low network constraint (i.e., brokers), while network constraint promotes attainment for those with low cultural fit.

Jun 2015 – Aug 2015 · 2 mos · Noida

• Worked in Marketplace Fulfillment and Logistics team of Paytm which is subjected to continuous modifications. I have written REST apis in Nodejs.

May 2015 – Jan 2016 · 8 mos

• I write this blog to help my juniors and fellow MS aspirants through the application process of Masters. Have received more than 27k views and still counting. I publish useful articles and answer to students' questions to guide them with the procedure. I believe students should be given proper guidance to make informed decision about their graduate course in the US and the universities and courses they choose, which is otherwise not easily available elsewhere.

Jun 2014 – Nov 2014 · 5 mos

• Research Project on "Analysis of anonymity of Bitcoin transactions and mixing services"

Dec 2013 – Jan 2014 · 1 mo · Noida

• Android Application Development for NFC (Near Field Communcation) testing at Test Innovation Group

Jul 2013 – Jul 2013 · 0 mo · Nodia

• Designed websites using ASP .NET and C#

Jun 2013 – Jul 2013 · 1 mo · Noida

• Created Android apps, worked on Java Native Development, Google Cloud Messaging

Jul 2012 – Jul 2012 · 0 mo

• Core Java