Everyone wants to work on self-driving cars or large language models. Nobody grows up dreaming about insurance pricing. And yet, after 18 years building ML systems for insurers, Luba Orlovsky would argue that insurance is quietly one of the hardest - and most interesting - problems in applied machine learning today.
The challenge is real: your model has to be accurate enough to keep the business solvent, explainable enough to satisfy a regulator, fair enough not to discriminate, and fast enough to price a policy in real time. There is no equivalent in most ML domains where all four constraints are non-negotiable at once.
And that was before climate change started invalidating historical risk data, electric vehicles broke every assumption built into motor models, and the EU AI Act arrived with binding requirements on algorithmic decision-making.
In this talk, Luba draws on real experience - from research lab to production systems running across 35+ countries — to explore what insurance reveals about the limits and possibilities of ML in the real world. The problems are specific to insurance. The lessons are not.