Soumi Sarkar

Leveraging autonomous AI agents and code-generation tools drastically accelerates feature development velocity, but it introduces a major structural risk. When AI models generate code across large, deeply integrated multi-module codebases, they frequently break state-hoisting principles, overlook complex lifecycle boundaries, introduce subtle memory leaks, or bypass established dependency injection scoping rules. Relying on human engineers to catch these micro-regressions during manual peer reviews slows down deployment pipelines, creating an operational bottleneck that defeats the purpose of automation.
Instead of forcing human reviewers to constantly police AI-generated pull requests, your application architecture should be intentionally engineered to enforce safety programmatically. Based on experience working directly within massive mobile monorepos like Microsoft Office (Word, Excel, PowerPoint), this session presents a platform-level blueprint for building a strict, compiler-driven verification harness that makes a large codebase structurally immune to automated generation errors.
Specifically, this talk will cover:
Deep Dive into Kotlin IR & Static Analysis: How to write custom lint checks and lightweight Kotlin Compiler Plugins to validate dependency graphs and block forbidden API patterns at compilation time.
Enforcing Architectural Contracts: Designing strict, type-safe interface boundaries and API isolation wrappers that visually and structurally restrict where AI agents can safely add code.
Building Autonomous Self-Correction Loops: Structuring a CI pipeline harness that parses compiler errors and failing unit test logs, automatically feeding them back to the generating agent for headless error mitigation.
Stop manually reviewing code differences and learn how to shift quality governance from human reviewers directly to the compiler.