I led development of natural language processing driven production features including Conversational AI for Virtual Agents, Incident Auto Resolution, Question-Answering for Search, and Text2Code for Workflows. We built next generation of Natural Language Processing technologies using some of the latest Language Models and Deep Learning techniques and scaling to thousands of large enterprise customers to improve work experience of 70M+ enterprise users.Conversational AI for Virtual Agents: (1) re-platformed PassageAI technologies and integrated with ServiceNow stack to bring deep learning based models to ServiceNow for Virtual Agents (VA). (2) created new enterprise language models known as ServiceNow Language models. (3) introduced new abstractions like Dialog Acts to make VA conversations flexible and natural. (4) TapeAgents: built a truly conversational experience using large language model for task oriented agents. (5) enabled 50X growth in the number of customers using conversational AI. Question-Answering for Search: (1) re-platformed PassageAI technologies to introduced Machine Reading Comprehension based Question-Answering in Search pipeline. (2) introduced Dense Passage Retrieval techniques for semantic search. (3) Large Language Models based Retrieval Augmented Generation (RAG) for answering questions.Text2Workflow : (1) created natural language description to workflows in ServiceNow flow designer using Code LLM; (2) created one of the first GenAI feature to recommend the next best action in ServiceNow flow designer using LLM.Incident Auto Resolution (IAR): created a new deep learning classification based pipeline for identifying intents from created incidents to automate workflows in Virtual Agents.Thought leadership: published multiple papers in the top conferences (e.g., ACL, IAAAI, EMNLP), multiple patents and external talks in Baylearn, Data Council, and other venues.

Built a conversational agents creation platform for customer and employee services using the latest Deep Learning and Natural Language Processing technologies. Led the NLP and Deep Learning efforts at Passage.AI. Built a world-class Machine Learning and Deep Learning team.Passage.AI's natural language understanding and processing platform can be used to create an intelligent conversational virtual agents for any website or business. These virtual agents can then be deployed with minimal effort on a website, mobile app, voice platforms such as Alexa or on messaging platforms such as Facebook Messenger and Slack. Passage AI got acquired by ServiceNow in February 2020.

I lead a very talented bunch of data scientists and engineers for building data driven products for the growth of network (such as People You May Know), membership and relationship. I work on intersection of products, machine learning and distributed systems for big data.People You May Know (PYMK): PYMK is a large-scale recommender system that analyzes billions of edges to predict social connections. Increased the number of connections from PYMK more than 15 times with new features and model improvement.Guests You May Know (GYMK): Built a new growth data product feature from scratch that is leading to significant membership growth, which extends PYMK's ability to bring new members on LinkedIn. Related Searches: Worked on search query recommender system that resulted in significant increase in the number of clicks for related searches through algorithm improvement. Also wrote and presented a research paper on technical details.Landing page optimization (Heathrow): worked on optimizing landing page after accepting a connection invitation that resulted in significant increase in the number of connections and new members joining LinkedIn.Feed Relevance: developed one of the first relevance based algorithm using PYMK generated connection strength.Thought leadership: Co-authored 8 papers related to recommender systems, link prediction, social network analysis, etc. Mentored three interns and co-mentored six other interns. Organized reading group for about a year and hosted multiple talks at Linkedin. Outreach through external talks and papers in: KDD, WWW, RecSys, CIKM, SIGIR, QCon, Web and Analytics Summit, Big data meetup, UT Austin, and USF. A blog post on Linkedin's main blog and three blog posts on LinkedIn's Engineering blogOther projects: Participated in eight hackday projects, winner of two hackday projects, and one hackday project on Linkedin labs so far. Three of the hackday projects incorporated in products.

I worked on building the next generation information retrieval platform. I worked on query, tweets and document classification/categorization, text analysis, query parsing, query expansion, information extraction, and search relevance. I also worked on building high performance, scalable systems for data fetching, data storage, and distributed processing.Tweet Classification: In the last project at Kosmix, developed a tweets classification system based on vector space model. This classification system was developed for Tweetbeat.com to classify incoming tweets from Twitter fire-hose to trending events. Built using Java, Perl, Python, and Tokyo Cabinet.Document Categorization: Worked on improving concept extraction from a given document (or tweet) to categorize the document among 10 million categories in the taxonomy. Also developed a framework to categorize hash-tags present in tweets. Built using C++, Perl, Ruby, Cassandra, and MapUpdate.Related Topics: In another recent project, developed a technique to analyze a given web page and figure out a set of related topics to explore. This technique is used by Kosmix's publisher solution product to provide links on web pages to explore related topics. Built using C++ and a taxonomy of 10 million nodes.

I worked on research and development of the next generation information retrieval platform. I worked on query parsing, query expansion, and search relevance to improve the quality of pages generated for any given topic. I also worked on building high performance, scalable systems for data fetching, data storage, and distributed processing.Topic Page: Developed the back-end for Kosmix Topic pages. Built the infrastructure to mashup results from various APIs and results from search index. Developed techniques to analyze and score (1) search results from various sources using techniques such as Okapi scoring, (2) the most relevant type of information for the given query, (3) the importance of a web page. Also, developed query parsing framework to find out location in the query, important concepts in the query, and to expand query to get more results related to the given query. Filed a patent for the novel techniques developed for relevance of topic pages. Built using C++ and Perl.Concept Extraction: Extracted 10 million music concepts (album, songs, artist) from MusicBrainz to add to our taxonomy of millions of nodes. Also created a module to show the music concepts on relevant topic pages. Built using C++, Python, Berkeley DB XML, and Mysql.Back-end Infrastructure: Developed a high performance, scalable HTTP thread pool using Curl library to asynchronously fetch web pages at run-time. This infrastructure is used by Kosmix topic pages to fetch more than 15 million web pages everyday. Also, developed a scalable caching framework to cache the whole topic page except advertisements and parts of the topic pages. Built using C++, Curl, and Ehcache.

I was a part of the Search Quality group, which is responsible for improving the quality of Google's search result. In particular, I was a part of the Automated Search Spam Filtering team. I researched and designed algorithms and heuristics for detecting spam in search results. I implemented my algorithms and performed experiments with a large amount (~ 13 TB) of crawled data, query logs, and toolbar data. I used more than 1000 machines to perform my experiments. Built using MapReduce, Sawzall, C++.

I was a part of the MSN Search group. I researched, designed, and developed a WML proxy server to empower mobile device users to search and surf the web through MSN Search. I performed query analysis experiments with a large amount (~ 200 GB) of query and crawled data to investigate machine allocations in a data center. Built using C#, ASP.Net.

I worked on IP Based Virtual Private Network (VPN), and extended this work into my senior undergraduate thesis.