Jun 2023 – Jun 2024 · 1 yr · San Mateo, California, United States · On-site

• As a part of the "Roblox Large Models" team at Roblox, whose mission is to shape the future of AI-driven game creation by leveraging cutting-edge machine learning techniques to develop multimodal large models that are capable of serving a diverse set of Roblox use cases such as content creation, better personalization, improved platform safety, etc. As a Principal Engineer in this team, my responsibilities include:
• 1. Design and implement machine learning algorithms that seamlessly operate across multimodal data such as text, images, audio, etc.
• 2. Take the lead in researching state-of-the-art techniques, stay at the forefront of multimodal machine learning, and contribute back to the community through publications and presentations.
• 3. Provide technical leadership for large/medium-sized projects of significant complexity and impact for Roblox, including coordinating project priorities and resources.
• 4. Undertake the critical task of processing multimodal datasets to leverage high-quality input datasets for model training and evaluation.
• 5. Leverage generative modeling techniques, including GANs (Generative Adversarial Networks) and VAEs (Variational Autoencoders) to create rich and expressive generative systems that produce novel, engaging, and aesthetically pleasing content.
• 6. Integrate Language Models for natural understanding capabilities, enabling the multimodal models to comprehend and generate content with enhanced linguistic richness.
• 7. Keep abreast of ethical considerations and ensure responsible AI practices in model development.

Feb 2023 – Jun 2023 · 4 mos

• 1. Leading research, prototyping and development of NLP based initiatives for Thoughtspot.
• 2. Tech leading natural language question answering using semantic signals and translation.
• 3. Tech leading building "chat gpt" kind of experience for Thoughtspot's in-app chat support.

Feb 2021 – Feb 2023 · 2 yrs

• 1. Managed a full-stack team of 6 developers responsible for improving indexing and search-quality for metadata discovery search of Thoughtspot.
• 2. Led a team of 5 to deliver autocomplete capabilities for metadata discovery search.
• 3. Helped drive product initiatives improve metadata-discovery-search click-rate of by 25% and number of searches by 50%.
• 4. Awarded "Operational-Excellence" award for Thoughtspot in year 2021.
• 5. Initiated the process of pre-mortems before any significant feature to ensure there is high likelihood of success once implementation is complete.
• 6. Created a culture of accountability and discipline in the team, where any major sev-0/sev-1 issue is followed up by a 5-whys style analysis to ensure learnings are prioritized.

Feb 2020 – Feb 2021 · 1 yr

• 1. Led prototyping of metadata-discovery search, which involved creating the first multi-tenant SaaS service for Thoughtspot, leveraging technologies like Kubernetes, Elasticsearch and Kafka.
• 2. Proposed and implemented the idea of exposing RED & USE metrics from each microservice and also involved in setting up the generic dashboard in grafana, usable for all microservices.
• 3. Led development of Answer Explorer feature, which involves generating one-click recommendations that help users get to insight-ful visualizations.
• 4. Proposed, POC'ed and executed on the idea of having graphql federation based API stitching in cloud for grpc based microservices using google-rejoiner.

Mar 2019 – Feb 2020 · 11 mos

• 1. Led a full-stack team of 5 engineers to implement the "Answer Explorer" feature - https://www.thoughtspot.com/press-releases/thoughtspot-advances-towards-autonomous-analytics-new-features-do-work-you
• 2. Led a team of 7 to help migrate Thoughtspot's internal build-system to compile C/C++/Python code to Bazel.
• 3. Drove design, setup and monitoring of remote-caching-server for bazel build-system, which reduced build-times across the company by 40% on average.
• 4. Drove reduction of memory footprint of usage-based-indexing service for Thoughtspot's proprietary search by 50% using pprof.

May 2018 – Mar 2019 · 10 mos

• 1. Part of 5 member team, which developed Natural-Language to SQL translation functionality from scratch to power Thoughtspot's NL/Voice input based search-bar.
• 2. This initiative involved creating the first of its kind "teachable" search experience, which was developed with a goal to bridge the gap of ~85% state of the art accuracy of NL to SQL translation against ~100% accuracy required in an enterprise setting.
• 3. Through this work this team was able to achieve multiple patent grants ( listed below in patents section).
• 4. This multi-year project was later productionized and termed "SearchIQ" ( see link below )

Mar 2016 – May 2018 · 2 yrs 2 mos

• 1. Part of a 2 member team, which helped create the data ingestion service for thoughtspot.
• 2. Wore multiple hats during my first few quarters in the company, where I pretty much worked across the stack ( AngularJS based SPA, Java middle-tier, Tooling, Core search functionality etc).
• 3. Part of 4 member team, which took on the challenge of translating natural-language query to SQL. This is an unsolved problem in the NLP research community. (State of the art accuracy is < 80% on most datasets).

Dec 2015 – Mar 2016 · 3 mos · Sunnyvale, CA

• Yahoo Gemini is the first unified marketplace for mobile search and native advertising. With Yahoo Gemini, advertisers get the performance and ease of search, combined with the scale and creativity of native advertising.
• Role:
• Develop techniques and algorithms for adselection; measure and optimize for relevance, click performance, and revenue.
• Gather and analyze data and identify problems, devise solutions, build prototypes, design and conduct offline simulation, offline editorial tests, and online experiments and buckets.
• Work closely with business and product teams in developing solutions and conducting experiments. Realize models and algorithms in products.

Oct 2015 – Dec 2015 · 2 mos · Sunnyvale, CA

• I work for the Video Search, Personalization and Analytics team. My resume (linked to Summary) has details about few projects that I worked on. Please reach out to me if you need further details or references.

Oct 2014 – Sep 2015 · 11 mos · Sunnyvale, CA

• I work for the Video Search, Personalization and Analytics team. My resume (linked to Summary) has details about few projects that I worked on. Please reach out to me if you need further details or references.

Jun 2014 – Oct 2014 · 4 mos · San Francisco Bay Area

Aug 2013 – May 2014 · 9 mos · State College, Pennsylvania Area

• Developed a simple privacy framework to provide suggestions to image privacy.
• Developed a multi-class/multi-label privacy framework to provide suggestions for image privacy. Images classified across three different actions and across 5 different privacy levels for each action.
• Used multiple computer vision techniques to encode image related features, which were used for learning classification models.
• Developed an ensemble based classification model to predict image privacy policies with an accuracy of 90%.
• Developed a new feature using the sentiment conveyed from an image to predict privacy levels. Sentiment in the context of image privacy was never studied till date.
• Code Repository - https://github.com/rmb347/imageWork

May 2013 – Jul 2013 · 2 mos · United States

• Designed a configuration format based on JSON for a new framework to develop payment processors.
• Developed an eclipse editor plugin for the configuration, which supported features such as Hyper linking, content completion, code generation, code folding etc.
• Developed payment processors for Mastercard and American Express which process approximately 10% of the total payment transactions by Amazon.

Oct 2012 – Feb 2013 · 4 mos · Survey Research Center

• I worked as an Android app developer. I was a part of developing the framework of an "Android Survey application", which the department used for conducting scientific research.

Aug 2010 – Jul 2012 · 1 yr 11 mos

• Designed and developed a generic persistence map which is disk-backed. It has the ability to recover data it stored after an application crash.
• Developed a custom memory allocator, which was designed to give out memory on demand. The model was developed to avoid call to heap manager when there is a memory requirement at runtime. This development was done considering cache coherency issues in NUMA.
• Developed and maintained nine Foreign Exchange (FX) trading applications, which were capable of trading around 700 million/week.

Jul 2009 – Dec 2009 · 5 mos

• At Broadcom, I worked for the enterprise networking team, for which I developed a tool called “Centralized Resource Server” which monitors and controls network activity of hardware devices. The application is used for resolving hardware utilization issues, scheduling tasks and sending alerts in case of a malfunction.