Machine Learning Guided Optimizations are already used in LLVM for decisions such as inlining and register allocation, but deploying trained models currently requires a custom compiler build with TensorFlow/XLA dependencies. This tutorial introduces the new MLIR-based path for embedding ML models in-tree: representing models with TOSA, lowering them through MLIR, and emitting C/C++ with EmitC. Using the inlining-for-size model as a running example, we will explain how the new flow separates training from inference, why it simplifies deployment across LLVM-based toolchains, and how to try the new tooling through a hands-on LLVM demonstration.