Agentic Engineering & ToolingConference50min
AI in the Dark: Running LLMs in Air-Gapped Production Environments
This talk describes moving an AI service from demo to production in an air-gapped defence environment with no internet, cloud APIs, or external services. It covers local LLM constraints, offline tool-calling, hallucination risks, and the need for human oversight in a fully isolated system.
talk.summaryAiDisclaimer
Mihaela Gheorghe-RomanSystematic
talkDetail.whenAndWhere
Thursday, October 8, 17:40-18:30
TBA 4
talks.roomOccupancytalks.noOccupancyInfo
Building an AI agent is easy. Building one you'd trust in production is harder. Building one that runs with no internet access, no cloud APIs, and no external services: that's a different problem entirely.
This talk follows a real journey: taking an AI-driven service from demo to production in an air-gapped defence environment where OpenAI, Azure and others are simply not an option.
The cloud disappearing changes everything: local LLMs under real performance constraints, tool-calling without external infrastructure, hallucination you can't fix by swapping an API, and human oversight in a system that talks to nothing outside its own network.
This is a production story from an environment where "it works on my machine" was never good enough and that's because the machine is isolated, locked down, and the usual shortcuts don't exist.
This talk follows a real journey: taking an AI-driven service from demo to production in an air-gapped defence environment where OpenAI, Azure and others are simply not an option.
The cloud disappearing changes everything: local LLMs under real performance constraints, tool-calling without external infrastructure, hallucination you can't fix by swapping an API, and human oversight in a system that talks to nothing outside its own network.
This is a production story from an environment where "it works on my machine" was never good enough and that's because the machine is isolated, locked down, and the usual shortcuts don't exist.
Mihaela Gheorghe-Roman
Mihaela is a Senior Software Architect at Systematic, Romania, and an Oracle ACE Associate for Java. PhD in Artificial Neural Network algorithms, Associate Lecturer at the University of Cybernetics and Informatics in Bucharest, she is passionate about building production-grade systems and bringing real-world experience to the stage.
In her professional career, spanning over 15 years, she has worked in the telecom and defence domains with a focus on Java, Spring Boot, C, C#, Scala and Angular, and a growing interest in AI systems and agentic architectures.
Outside of tech, she holds a black belt in traditional Jiu-Jitsu, runs marathons, and travels the world whenever she gets the chance.
In her professional career, spanning over 15 years, she has worked in the telecom and defence domains with a focus on Java, Spring Boot, C, C#, Scala and Angular, and a growing interest in AI systems and agentic architectures.
Outside of tech, she holds a black belt in traditional Jiu-Jitsu, runs marathons, and travels the world whenever she gets the chance.