This talk introduces Modular MAX’s **MLIR-based Graph Compiler**, a system designed to overcome the difficulty of extending operator sets in traditional deep learning compilers. Rather than requiring complex rebuilds or separate DSLs, MAX's framework uses a pragmatic, "kernel-first" approach built for extensibility. At its core is a JIT compiler that optimizes model graphs for specific hardware. The key innovation is its use of **Mojo primitives**—such as decorators, parameters, and traits—that empower developers to easily register custom, high-performance kernels directly, without needing to modify and rebuild the underlying compiler infrastructure. The presentation will walk through the compiler's phases, focusing on this extensibility and demonstrating how Mojo-authored kernels interact with the system.