Staff Research Lead, ServiceNow Research · Adjunct Professor, Mila · Montréal
I work on the data, benchmarks, and executable environments that decide what AI agents are trained and measured on, across documents, screens, terminals, and long-horizon enterprise workflows. Recent work includes EnterpriseOps-Gym (ICML 2026), a stateful benchmark for agentic planning and tool use in enterprise systems, and UI-Vision (ICML 2025), a desktop GUI benchmark for visual perception and interaction.
My working premise is that evaluation directs optimization: whoever defines what counts as hard, and what counts as solved, sets the training agenda for everyone downstream. So I focus on the instruments that steer capability: benchmarks that isolate real failure modes, datasets grounded in genuine human demonstrations, and environments recovered from real systems where agents can act, fail, adapt, and be scored.
I completed my Ph.D. at Mila / Université de Montréal under Aaron Courville, and spent time as a Research Scientist Intern at Google DeepMind. Before that, I studied computer science at IIT Delhi.
What I build, at a glance
Stateful, policy-governed systems of record where agents act through tools and are graded on the final state of the world.
Measuring whether agents can see and act on real screens: desktops, web apps and professional software.
Licensed, human-grounded data at scale: the raw material for training and measuring multimodal agents.
Small, specialized models and simple agent interfaces that outperform larger general systems on the right data.
Agents that imagine before they act, from pixels and robots to enterprise systems.
Venues shown are for the main publication; in-progress items are ongoing work.
Collaboration: I'm excited to work with students, interns and collaborators on enterprise agents, terminal and computer-use agents, digital-world simulation, and world models for reliable intelligence. Get in touch.