This new DSL infrastructure addresses the limitations of traditional methods for generating MLIR from Python, which often sacrifice Python's dynamic nature or fail to capture control flow correctly. The proposed solution introduces a **hybrid model** that combines AST-based rewriting with runtime tracing, enabling seamless interoperability with native Python libraries while ensuring robust IR generation. A second key innovation is its **target-dialect-agnostic design**, which allows developers to build new DSLs for various MLIR backends by extending a shared base abstraction. This framework has been successfully demonstrated through the **CuTe DSL**, a Python frontend for NVIDIA’s CUTLASS library, which is fully open-source.