Roy Saurabh

Roy Saurabh is a researcher working at the boundary between AI systems and the institutions that govern their use. His current work is on AI assurance: building reproducible methods and software that make evidence about AI systems auditable rather than asserted, with particular attention to privacy-preserving and federated approaches where data cannot be centralised. He holds a doctorate in mathematics and computer science from Université Paris Cité, defended in December 2024, on machine learning from wearable data for mental-health monitoring of healthcare professionals. Before returning to full-time research he spent seven years at UNESCO, working on learning technology, teacher development and education data governance, and was part of the UNESCO writing team for the Broadband Commission's report on data for learning. He serves on UNICEF Innocenti's expert advisory group on the governance of children's data, and is an advisor to the JA Institute.

• AI assurance and accountability, Claims about an AI system should be checkable by someone who was not involved in building it.
• Privacy-preserving collaborative AI, Institutions that cannot pool sensitive data can still learn together, if the governance is built alongside the method.
• Human-centred adaptive systems, Systems that adapt to people need to be honest about how little they know about the person in front of them.
• Data governance for public-interest systems, The governance of education and children's data decides what the technology is allowed to become.

1. Roy, S. (2026). BA-FedSHAP: A reproducible toolkit for auditing background-induced attribution drift [Preprint]. SSRN. https://doi.org/10.2139/ssrn.6864992
2. Roy, S. (2024). Data preprocessing and machine learning in wearable data analysis: Assessing efficacy and challenges for mental health monitoring of healthcare professionals [Doctoral dissertation, Université Paris Cité]. https://doi.org/10.70675/bdabcabdz273cz4ff3z8b79zc9434acdf00d
3. Roy, S., Singh, N. C., & Duraiappah, A. K. (2019). The alternate education for the 21st century. In 2019 IEEE 18th International Conference on Cognitive Informatics & Cognitive Computing (ICCI*CC) (pp. 265-271). IEEE. https://doi.org/10.1109/ICCICC46617.2019.9146061

1. Roy, S. (2026). CSA-Lite: Context-Sliced AI Assurance Lite (Version 0.2.2) [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.20403165
2. Roy, S. (2026). LEXON-Bench: Executable AI regulatory obligation reasoning [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.20397383
3. Roy, S. (2026). RCEA Passport Engine: Role-conditioned evidentiary adequacy reference implementation (Version 0.1.1) [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.20384669
4. Roy, S. (2026). Provena: Reproducibility package for evidence-laundering stress tests [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.20376679
5. Roy, S. (2026). BA-FedSHAP: A reproducible toolkit for auditing background-induced attribution drift [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.20356218

1. AI in health: how smart technology is breaking new frontiers in medical care (2026). European Commission (DG CNECT). Role: Author.
2. Federated learning for children's data: encoding trust in education systems (2025). UNICEF Innocenti. Role: Author.
3. The Transformative Potential of Data for Learning (2023). ITU/UNESCO. Role: Writing team.
4. The Transformative Potential of Data for Learning, interim report (2022). ITU/UNESCO. Role: Writing team.
5. CHI, a knowledge-sharing digital platform powered by machine learning (2019). UNESCO MGIEP. Role: Author.