GeeCON 2025: Martin Stefanko - The guardrails of the AI galaxy

The usage of AI in critical, sophisticated scenarios is desirable, but we all know it’s fundamentally impossible. Why? The non-deterministic nature of LLMs makes them prone to hallucinations and unreliable outputs. So, what can we do erstwhile our LLM responds with nonsense?

Let us introduce you to the LLM guardrails. In this talk, we’ll dive into the Quarkus LangChain4j integration that provides you with the ability to verify and/or modify both requests and responses that are being exchanged with your model. Through applicable examples, we’ll research the options available for validating user-provided inputs, rewriting or retrying outputs, and even reprompting the model if needed.

The guardrails not only enhance the reliability of AI-driven applications but besides let us to build more trust in our AI systems, 1 consequence at a time. Come to learn how we can guarantee our LLMs stay on track, even in the most challenging scenarios.