Building Ontology Guardrails for Agentic AI

In this 90 minute hands-on session, you’ll build the guardrail layer from the keynote with your own hands — no prior knowledge of ontologies required, just the ability to read and edit a little code. Working in a browser-based notebook (nothing to install), you’ll start by writing simple triples to model a real business domain, then declare RDFS and OWL relationships and watch a reasoner infer facts you never explicitly stated. From there you’ll turn to SHACL, the constraint language that does the opposite job — enforcing rules and rejecting data that violates them — and you’ll write your own shape that catches a bad, plausible-looking action before it slips through. You’ll leave understanding, concretely, why probabilistic models need a logical layer outside them, and how the pieces — triples, inference, and validation — fit together into a working guardrail. Time permitting, we’ll close by dropping one of your SHACL shapes into a live agentic loop, so you can see the guardrail stop a bad tool call in real time — and take the notebook home to keep building.