Working with LLMs (large language models) for applications in Bing Chat and Business Chat.

I am working on research for conversational assistants at Semantic Machines, part of Microsoft.- Built an early prototype for Microsoft Business Chat, as part of a team with 5–10 engineers and researchers. Conceived an initial evaluation plan. Designed a framework for real-time, asynchronous client–server sync.- Developed methodology and infrastructure for leveraging large language models in semantic parsing by paraphrasing and grammar-constrained decoding, to enhance generalization new domains and APIs.- Mentored interns on projects relating to model compression in NLP, user privacy, and crowdsourcing. Helped formulate research questions, select benchmark datasets, and design empirical evaluations.- Contributed to internal infrastructure for unit testing, code review, and open sourcing of research code.- Collaborated with teams across Microsoft for sharing best practices on large language model applications

Intelligent Systems Lab, directed by Vladlen KoltunResearch in neural program synthesis

Microsoft Research AIContinuation of neural program synthesis research from summer internship.

Microsoft Research AIResearch in neural program synthesis.

Research in neural program synthesis.

Google Brain/Google ResearchResearch related to automated theorem proving and generative modeling of structured data.

Google TranslateNeural machine translation

Parallel document mining for Google Translate.I worked on improving a system for finding web documents that are translations of each other. These document pairs are then used for training Google Translate's statistical machine translation back-end.

With Professor Dawn Song's research group.Smartphone security- Developed a translator from Dalvik to JVM bytecode, to analyze Android apps with existing Java-based tools. - Built a tool to automatically exploit vulnerable in-app billing implementations in Android applications. Web security - Discovered vulnerabilities in Facebook and Google’s (mis-)use of HTML5 cross-document messaging.- Investigated the prevention of cross-site scripting vulnerabilities with automatic sanitization of untrusted data. - Identified examples of client-side JavaScript input validation in production web sites. Miscellaneous - Contributed to authorship identification experiments on blogs using statistical machine learning techniques. - Implemented a SMTP client in Python for a project to automatically infer network protocols.

Graded and helped create homework assignments, projects, and exams for Discrete Math, Machine Structures, Artificial Intelligence, and Algorithms courses under supervision of different professors.

YouTube eCommerceWorked as part of team that launched http://www.youtube.com/movies. Implemented short-to-long video links, and tools for migration to more flexible backend data format. Prototyped new subscription payment model for bundles of paid videos.