Architecture & ModernizationTools-in-Action30min
Stopping AI-Generated Architectural Drift: Deterministic Guardrails for Agentic Coding
This talk explains how to prevent AI-generated code from spreading architectural violations by using layered guardrails. It shows why natural-language rules fail, and how deterministic enforcement with ArchUnit, custom linters, and CI gates can stop bad patterns immediately and make higher-level rules effective.
talk.summaryAiDisclaimer
Sasha PodlesniukINSHUR
talkDetail.whenAndWhere
Monday, October 5, 17:35-18:05
TBA 3
talks.roomOccupancytalks.noOccupancyInfo
Agentic coding generates code faster than any team can review it, and that speed lets a single questionable pattern spread across a codebase before anyone notices. This talk shows how to stop that drift at the source, with layered guardrails that make architectural violations structurally hard to introduce instead of something you catch too late in review.
Drawn from reviewing hundreds of AI-assisted pull requests: natural-language (NL) repo rules don't hold the line against an LLM. The reliable fix is a deterministic foundation, executable architectural rules acting as fitness functions for AI-generated code, built with ArchUnit, custom linters, and CI gates that fail the build the instant a boundary is crossed. That layer is also what makes NL rules work: once it fences off the cases that must never be wrong, targeted NL rules become precise instead of suggestions an agent ignores. Includes live demos of guardrails catching violations in real time.
Takeaways: why NL rules fail to stop LLMs propagating bad patterns; how to build a deterministic enforcement layer (ArchUnit, custom linters, CI gates) with instant feedback; how that layer makes your NL rules reliable; a practical model for sorting any rule into the right layer.
Audience: software architects, tech leads, and senior engineers maintaining architectural consistency on AI-augmented teams.
Drawn from reviewing hundreds of AI-assisted pull requests: natural-language (NL) repo rules don't hold the line against an LLM. The reliable fix is a deterministic foundation, executable architectural rules acting as fitness functions for AI-generated code, built with ArchUnit, custom linters, and CI gates that fail the build the instant a boundary is crossed. That layer is also what makes NL rules work: once it fences off the cases that must never be wrong, targeted NL rules become precise instead of suggestions an agent ignores. Includes live demos of guardrails catching violations in real time.
Takeaways: why NL rules fail to stop LLMs propagating bad patterns; how to build a deterministic enforcement layer (ArchUnit, custom linters, CI gates) with instant feedback; how that layer makes your NL rules reliable; a practical model for sorting any rule into the right layer.
Audience: software architects, tech leads, and senior engineers maintaining architectural consistency on AI-augmented teams.
Sasha Podlesniuk
I am a Domain Tech Lead at INSHUR with more than 12 years of experience working on digital transformation initiatives across fintech, insurtech, and cybersecurity. My work has included modernizing legacy platforms, migrating monoliths to microservices, and building large-scale cloud-native systems.
More recently, I have focused on applying AI to digital transformation, leading the development of agentic AI systems and integrating large language models into operational workflows. I am interested in making AI and complex technology initiatives practical, accessible, and understandable across engineering, operations, and business teams.
More recently, I have focused on applying AI to digital transformation, leading the development of agentic AI systems and integrating large language models into operational workflows. I am interested in making AI and complex technology initiatives practical, accessible, and understandable across engineering, operations, and business teams.