Abstract Machine Learning (ML) inference is shifting from using pre-developed static, CUDA C++, GPU kernel libraries to using MLIR-based graph compilers that perform advanced optimizations and generate custom kernels. This paradigm shift reimagines how we achieve ML inference in safety-critical domains such as automotive applications. Traditional approaches relied on qualifying static kernel libraries—pre-built for fixed input shapes and parameter ranges—according to the ISO 26262 standard. However, the demanding performance requirements of diverse ML models and rapidly evolving hardware accelerators necessitate generating optimized kernels on the fly, which only ML graph compilers can provide. This paper presents an industrial experience report on a comprehensive verification framework for ML inference in automotive applications. We describe the transition from static kernels to dynamic ML graph compilation and introduce two complementary verification strategies: (1) formal methods targeting memory safety and concurrency properties in CUDA kernels and MLIR-based compiler; and (2) AI-driven testing for functional correctness. Our experience over multiple years of production use demonstrates that validating ML graph compiler output can satisfy the ISO 26262 ASIL B requirements - without requiring compiler tool qualification - while enabling performance and flexibility benefits. We discuss remaining challenges including scalability of formal verification and adapting to evolving compilers and hardware platforms.