I am an AI safety and robotics researcher working on reliable evaluation, runtime safeguards, interpretability, and trustworthy machine learning systems. My research studies how intelligent systems fail under real-world conditions and how those failures can be measured, monitored, and controlled. I am particularly interested in agent safety, vision-language-action models, distribution shift, and statistical methods for dependable AI evaluation. I completed my master’s degree in Data Science at New York University’s Center for Data Science. Before founding and directing LatentOps, I was Co-founder and CTO at Offtrack.
2026
Published ProxyGuard, a framework for reliable inference over randomized data-release mechanisms.2025
Introduced TriGuard for evaluating robustness, attribution entropy, and explanation drift.I separate what an SAE coordinate encodes from how the adapted model uses it, showing that semantic selectivity can persist across fine-tuning even as downstream sensitivity and intervention effects change.
I study how the access conditions around dual-use assistants change both legitimate utility and harmful actionability, treating safeguard evaluation as a deployment-level measurement problem.
We test whether OpenVLA’s internal activations contain useful warning signals before task failure when visual conditions shift, and examine where retrospective discrimination falls short of deployable monitoring.
We explore sparse explicit key-value memory inside pretrained language models as an alternative to repeatedly placing retrieved documents in context, with the goal of improving reasoning efficiency.
We examine why common flatness and sharpness measures can be misleading in deep networks, where reparameterization and model symmetries can change loss geometry without changing predictions.
01
Runtime AI safety — monitoring and controlling risk when autonomous systems use tools, data, and external services.02
Reliable evaluation — designing measurements and statistical guarantees that remain meaningful after selection and distribution shift.03
Robotics and interpretability — studying internal signals that reveal when embodied models are likely to fail.2026
Area Chair, Who Verifies the Agents? Toward Reliable Agent Development, NeurIPS 2026 Workshop.Maintaining
Pari — a Rust-powered similarity indexing and deduplication engine with Python bindings, a CLI, and persistent or shared backends.Contributing
Ordin — local command intelligence and pre-execution safety for developers and AI agents.