Research LLM interpretability/control/fine tuningLLM safety fine tuning: Representation Tuning (https://arxiv.org/pdf/2409.06927; accepted to NeurIPS 2024 MINT workshop)LLM self-recognition ability: Inspection and Control of Self-Generated-Text Recognition Ability in Llama3-8b-Instruct (https://arxiv.org/pdf/2410.02064)Affiliations:BlueDot Impact, AI Safety Fundamentals, Alignment June 2024-PresentCourse Facilitator/Teaching FellowSupervised Program for Alignment Research (SPAR) June 2024-Sep 2024Researcher

Lead Quantitative User Experience Researcher, Google Discover Ads August 2021-PresentCreate, evangelize, and support an LLM-based system for automatic clustering and summarization of free-text survey responses, utilizing prompt engineering/iterated prompting with Gemini and PaLM models, which has been adopted by researchers company-wide.Assess user ad experience through surveys and logs metrics to support business-critical ad inventory expansion driving XXX% revenue growth.Identify opportunities for auction and targeting improvements through logs analysis and statistical modeling.Develop ad user value survey scale to inform format and auction experiments and platform health.July 2016-July 2021: Google for EducationGuide the development of Google's educational technology offerings through product and market research, statistical data analysis, and predictive modeling. Work with PM, UX, Eng, Sales, and Marketing leads to identify and prioritize current research needs and initiate foundational research in strategic opportunity areas. Design and implement metrics and evaluation criteria. Scope spans domains (K12, higher ed, international, etc), products (Google Classroom/GSuite, ChromeOS, Search, YouTube, i.a.), and technology areas (LMS, online learning, personalized learning, i.a.).Jan 2019-April 2019: Google.org FellowWorked with Goodwill as part of a team of Google.org Fellows to build out their data collection and analytics capabilities.Feb 2012-July 2016: Google AdWordsLed a team of quantitative user experience researchers focused on pursuing high-impact research opportunities on Google’s key revenue-generating product.

Within the Natural Language Processing group, I helped develop a statistical system for automatic document clustering based on named entities, conducted research on psycholinguistically and neurolinguistically plausible models of syntax, and developed a software tool to assist users in Arabic name transliteration. Within the Robotics group, I developed a prototype visual navigation reinforcement learning algorithm.

Developed biologically motivated vision algorithms for scene classification autonomous robot navigation.

Developed natural language processing applications at the Information Sciences Institute: worked on a project to establish quantitative metrics for text coherence and a statistical system for automatic rewriting for greater text coherence.

Developed on-line reporting software to query databases and compute and display statistical survey results.

Worked on natural language processing product used to answer customer questions on web sites. Developed web-based reporting tool to analyze customer queries and quality of responses and display results.

Developed distributed, web-based applications for the US Postal Service. Projects focused on exposing USPS functions such as rate calculation, package tracking, and label generation to online businesses over the internet. Created online helpdesk sites used by support center staff to view and update user records and demonstration e-commerce site used to illustrate incorporation of APIs.