Frederic Rivain

Most AI deployment assumes you can see the data you're working with. But what do you do when you cannot? When regulation, contractual commitments, or a zero-knowledge architecture mean the model must deliver intelligence without ever accessing raw user content? This session presents a concrete engineering approach to that problem, drawn from building AI on top of a credential management system where all customer data is encrypted end-to-end. We cover the full lifecycle: designing privacy boundaries before training begins; building training data sets and evaluation loops that do not rely on customer data; detecting distribution gaps without exposing sensitive inputs; improving models when debugging cannot involve raw data; and managing drift in production. We then examine the infrastructure layer: how cloud secure enclaves and on-device inference each map to different risk profiles, latency requirements, and threat models. Attendees will leave with a playbook for treating privacy as a structural constraint from day one, and a framework for choosing the right inference infrastructure for systems that cannot afford to expose customer data.