AI researcher and engineering leader with internationally recognized contributions. Nearly two decades building production ML systems across academic research, startups, and big tech. 120+ publications and patents (14,000+ citations, h-index 32). Co-author on Gemini 1.5, Gemini 2.5, and Gecko. Led teams of 60+ scientists. Pioneer of sentence simplification and distributional semantics methods for biomedical NLP.
Principal Machine Learning Engineer
2026 - Present
- Leading PINSAFE, a co-evolutionary multi-agent safety framework for generative AI: 7-dimension taxonomy (52 harm types, grounded in GDPR/OWASP/NIST), four specialized agents (red teaming, evaluation, filtering, monitoring), and the methods for co-evolving them that are foundational to safety at Pinterest.
- Designing PinJev, a multimodal classification model for image+text safety, ads, and content mining.
- Building PinValue, a feedback-loop system for content acquisition improving relevance from ~5% baseline.
AI Researcher
2025 - 2026
- Led AI and modeling for on-device agentic multimodal AI on Meta AI glasses (Ray-Ban Meta): designed the on-device model and agentic system powering hands-free message condensation, email recall, and contextual query answering.
- Shipped messaging intelligence: automatic announcement, group chat summarization, specific message recall ("What did Jamie suggest for dinner?"), and visual/verbal memory systems with review/edit in the Meta AI app.
Senior Staff Scientist / Manager
2022 - 2025
- Contributed to safety of Core Gemini model through instruction following, safety heads, and RL. Led team that reduced overblocking by 50%.
- Led PaLM Embeddings powering Vertex AI and PaLM API. Co-authored Gecko (256-dim outperforms all 768-dim on MTEB).
- Powered retrieval for NotebookLM and Gemini API's Semantic Retriever with SOTA embeddings.
- Collaborated with AstroHelper/Gmail on LLM integration for email and scheduling; with OCTO on talk-to-manuals for Ford and GM.
- Built first voice order-taking demo applauded by Wendy's executives and featured in WSJ.
Principal Applied Scientist, Alexa AI
2020 - 2022
- Led science for Natural Turn-Taking (Conversation mode, CNN). Authored two submitted patents.
- Mentored 60+ on Alexa Conversations. Advised Stanford, UPM for Alexa Prize SGC4.
- Initiated funded collaborations with faculty from Columbia, UC Davis, UT Austin, and Emory.
- First author of three-year innovation roadmap presented by Rohit Prasad across Alexa.
Chief Scientist and SVP
2017 - 2020
- Led Conversational AI for Intelligent Virtual Assistants serving 1,500+ enterprise customers (CenturyLink, IBM, Mercedes-Benz, Sacramento Kings).
- Architected deep learning with pretrained transformers fine-tuned on 30M+ responses for NLU, policy decisions, and RL-augmented NLG.
- Streamlined HitL to handle exceptions within 10-second SLA; designed self-learning XR pipelines and annotation microservices.
Guest Faculty, "NLP with Deep Learning"
2018 - 2021
Principal Applied Scientist, Cortana
2016 - 2017
- Led conversational understanding across Cortana, Office 365, LUIS, and external products. Mentored 15+ in Bellevue, Redmond, and China.
- Led design of AI efforts improving precision, recall, and automation for intent and entity understanding.
- Delivered results central for winning key demos to Bill Gates, Harry Shum, and David Ku.
Assistant Professor / Adjunct Professor
2013 - 2018
- Led 25+ affiliate members with 15+ grants from NIH, Novartis, AbbVie, Baxter, Mayo Clinic, and Stanford.
- Built NLP for clinical trial enrollment (used by Novartis and Northwestern Medicine), EHR mining, cardiovascular phenotyping, and echocardiography report extraction (EchoInfer).
- 2nd place in TREC Clinical Decision Support track. Reviewed grants for NIH. PC member for 15+ AI conferences.
Principal Investigator
2011 - 2013
- Supervised 5 junior PhDs and interns as PI of three grants. Operationalized cTAKES in Enterprise Data Warehouse.
- Built AskMayoExpert NLP system for clinical QA. Developed coreference resolution system for clinical narratives.