Defence-in-Depth: How We Build Security for Diia.AI [ukr]

What happens when a large language model becomes the entry point to government services that operate under real-world load and in the context of an information war? In such an architecture, any request may be not only incorrect but also intentionally manipulative — and standard AI safety solutions prove far less reliable than they appear in laboratory benchmarks.

In this talk, I will share how we built a custom guardrail module for Diia.AI after encountering the limitations of off-the-shelf filters and the high cost of the LLM-as-a-Judge approach. Instead of validating every request with a large model, we designed a cascade security architecture: fast ML classifiers filter out most of the traffic, while the LLM is invoked only where deeper contextual analysis is truly required.

This talk is not about perfect models, but about trade-offs, constraints, and practical decisions that must be made when an AI system operates not on a laptop, but within a national-scale service.