Smarter email infrastructure configuration: LLM assistant for Halon’s HSL

Large Language Models (LLMs) offer powerful new capabilities for interacting with and understanding complex systems. In this presentation, we share our experience building an LLM-powered assistant tailored to a Domain Specific Language (DSL) used for configuring Halon, a modern email infrastructure. This DSL encapsulates intricate routing logic, security policies, and protocol-specific behaviours—making correctness critical and manual configuration error-prone.

We explore how we adapted a general-purpose LLM to understand the structure and semantics of HSL (Halon Scripting Language), combining prompt engineering and integration with validation tools. The assistant is integrated directly into the Visual Studio Code Integrated Development Environment (IDE), enabling users to author, analyse, and troubleshoot configurations more efficiently through context-aware suggestions, explanations, and error detection. We discuss challenges such as disambiguating user intent, maintaining DSL syntax constraints, and ensuring the assistant's outputs are safe for production environments.

Through real-world usage data and examples, we highlight the assistant’s impact on developer productivity and configuration reliability. This talk will be of interest to anyone working at the intersection of AI and system configuration, especially in domains where correctness and security are paramount.

Bio: Carlos Pereira is a Principal Machine Learning Engineer at Halon, a leading provider of email infrastructure for service providers. He brings over 14 years of experience as a Software Engineer, with more than 7 years dedicated to applying machine learning in production systems. Carlos is a co-inventor of granted patents in both the US and the UK, and holds a Master’s degree in Software Engineering.