Roki Seydi

Through line

AI systems encounter people through data — records, clinical notes, interaction logs. The assumptions embedded in those representations about who people are, how they behave, and what they need are almost never tested against humans as they actually are.

The gap

Labs now ship domain products in finance, healthcare, and life sciences on the same general model: the same underlying system, wrapped in a domain prompt and evaluated on benchmarks scoped to that vertical. A 2026 industry benchmark found four of six frontier models fabricated financial data on loan documents with realistic missing information, two of them confidently and without disclosure. A Mount Sinai study found 52% undertriage of emergencies in the consumer ChatGPT product, against OpenAI's own physician-benchmarked safety claims. Life sciences is scaling at the same pace — including a $200M, four-year Anthropic and Gates Foundation commitment targeting 4.6 billion people in global health — with almost no comparable independent evidence. The pattern holds: strong published performance under test conditions, almost none under real deployment.

Practice

I am building an independent evaluation practice for that gap — auditing frontier labs' vertical-specific deployments against their own performance claims. Testing a deployed system is not optional; it is the condition of deployment. The evaluation methodology runs from computational patient models and social agents that simulate heterogeneous real-world contexts, through context-grounded and adversarial testing, to monitoring and audit infrastructure for agents operating at scale in regulated environments.

My work

Healthcare is my primary domain and the first vertical I am stress-testing systematically. My work identified cultural confabulation — a previously uncharacterised failure mode in frontier model reasoning for healthcare — and built open-source evaluation infrastructure on the UK AI Safety Institute's Inspect platform that goes beyond existing benchmarks to measure systematic deployment failures across frontier models.

Background

I am an inaugural Spärck AI Scholar at University College London, where I am completing an MRes in AI-Enabled Healthcare and am a member of the Health Algorithm Lab. I hold an MPH from Imperial College London and a BA from King's College London. I have presented at the Technical AI Safety Conference at Oxford, moderated at the Skoll World Forum, and review for ICML and NeurIPS.

Otherwise

I went to art school for five years and played in orchestras for nearly a decade. I grew up by Lake Como, Italy.