Large Language Models (LLMs) enable powerful conversational systems, but they also introduce a key challenge: hallucinations. In real-world applications, this can lead to outputs that are not grounded in system capabilities or user input, reducing reliability and trust.
In this roundtable, we present practical techniques used in production to improve the reliability of LLM-driven systems. By introducing structured reasoning fields, careful response schema design, and hybrid validation that combines LLM reasoning with system-side checks, we guide models to analyze constraints before producing actionable outputs.
These techniques significantly reduced incorrect responses while preserving the model’s ability to correctly handle valid requests. The session shares practical patterns for building more reliable and controllable LLM-powered applications in production environments.
Shir Hilel is an ML Engineer at Vonage with over 10 years of experience delivering various production-grade solutions, ranging from classic ML to LLM-powered systems, with a strong emphasis on measurable, data-driven outcomes. She holds both a Bachelor’s and a Master’s degree in Information Technology from Bar-Ilan University.
Keynote session: Hadas Grossmon Ella
Break
Lightning talks session
Roundtable closing
Talk by Hila Paz
Talk by Dr. Moran Mizrahi
Closing remarks
End
