Post-Training & Agentic Large Language Models at Meta
Building the next generation of LLMs through reinforcement learning and post-training, inducing agentic reasoning, tool-use, and factuality in foundational models deployed to hundreds of millions of people worldwide.
I am a Senior Research Engineer at Meta, where I explore post-training research on foundational language models. My work focuses on inducing and enhancing agentic capabilities in LLMs through reinforcement learning, advancing how AI systems reason, plan, and act across multi-turn, multi-tool, and multi-modal environments.
Beyond research, I have shipped production AI features for Wearables at scale. I contributed to the launch of Ray-Ban Meta smart glasses, Ray-Ban Meta Displays, and Oakley sport glasses, delivering real-time sports experiences across major sporting events and bringing Meta's Fitness AI features to millions of users worldwide.
Prior to Meta, at VMware I co-led the development of commercially viable open-source instruction-following LLMs, now downloaded tens of thousands of times monthly on Hugging Face. I also architected the LLM API infrastructure that serves over 2,000 daily requests across 40+ internal teams.
I hold an M.S. in Computer Science from the University of Illinois at Chicago. My research has been published at ICML, EMNLP, and IEEE, and featured in IEEE Spectrum, The Register, and Business Insider. I am committed to pushing the frontier of AI.
Co-established VMware's Hugging Face organization. Published 15 instruction-tuned LLMs and 2 NLP datasets, downloaded 243K+ times by researchers, engineers, and enterprises worldwide.
Peer-reviewed publications at ICML, EMNLP, and IEEE, plus technical articles and open-source research. View full profile on Google Scholar.
Introduces a reinforcement learning framework that directly optimizes LLM truthfulness by reducing hallucinations by 28.9% and improving truthfulness by 21.1%, while enabling models to abstain when uncertain. Currently deployed in production at Meta Reality Labs, addressing critical AI safety priorities.
Proposes an efficient fine-tuning framework that improves RAG factuality by 5.4% by training models to handle distractor passages and plan reasoning strategies, now deployed in production systems.
Demonstrates that AI models automatically generate more effective prompts than human engineers through systematic optimization, discovering unexpected high-performing strategies (e.g., Star Trek roleplay) that boost mathematical reasoning in LLMs. Featured in IEEE Spectrum, New Scientist, and 17+ major publications, reaching 500K+ engineering professionals.
Develops an efficient CNN-based face detection model achieving 72% mAP on challenging multi-scale, multi-angle, and occluded face scenarios, running at 60fps on consumer GPUs for real-time applications.
Peer review and program committee service at international AI conferences.
Reviewed 10+ manuscripts across top-tier venues, evaluating research on AI safety, hallucination reduction, and machine learning systems.
Coverage of my research and expert perspectives in leading technology publications, and more.
Open to conversations about frontier AI research, whether collaborations, advisory roles, or the right next challenge.