Ask a cat to do something predictable. Now ask an LLM (large language model). Congratulations—you've discovered the same challenge.
We've all tested the limits as users: the off-topic question, the creative workaround, the perfectly reasonable request that gets a wrong answer. But as engineers, we're on the other side—trying to anticipate every unexpected interaction and every unpredictable response. It's like herding cats, except the cats can generate novel content and your users are actively teaching them new tricks. The challenge isn't just making these systems work—it's making them safe, compliant, and trustworthy.
‍
This talk explores practical strategies for bringing order to chaos. We'll cover what guardrails are, examine different types and implementation approaches, analyse their trade-offs, and discuss how to test systems designed to surprise you. By the end, you'll have actionable knowledge for building safer systems—even when herding cats would be easier.