Roy Saurabh

I am a researcher working at the boundary between AI systems and the institutions that govern their use. My work asks a narrow question with wide consequences: what has to be true, technically, and institutionally, before an AI system can be trusted in a setting where it affects people's lives.

That question has three parts, and I have spent my career moving between them rather than settling in one. There is what we can know about a system: how it behaves, on whose data, with what uncertainty. There is what institutions should require before they rely on it: what evidence is adequate, for which decision, held by whom. And there is how those requirements become operational, the software, the audits and the governance machinery that turn a principle into something a regulator or a hospital can actually check.

My doctoral research, defended at Université Paris Cité in December 2024, sat in the first of those parts. I worked on machine learning from wearable sensor data to monitor the mental health of frontline healthcare professionals, a problem where the data is sparse, intimate and consequential, and where the honest finding is as often about what the method cannot support as what it can.

Before returning to full-time research I spent seven years at UNESCO, in Paris and previously at its institute in New Delhi, working on learning technology, teacher development and the data governance of education systems. That period taught me something research alone does not: how evidence actually travels into policy, how much of it is lost on the way, and how often a well-founded technical caveat disappears by the time a recommendation is drafted.

I now work on AI assurance, building tools that make evidence about AI systems auditable rather than asserted. This includes reproducible frameworks for deployment-conditioned risk analysis, executable reasoning over regulatory obligations, and methods for testing whether audit evidence is adequate for the specific decision it is being used to justify. I also serve on UNICEF Innocenti's expert advisory group on the governance of children's data, which keeps the work close to a population that bears real cost when data governance fails.

I write here about that work, and about the parts of it that do not fit neatly into papers.