Worked on the iree-org/iree repository to enhance the Flow dialect by enabling device-affinity global replication through a controlled experimental rollout. This involved relocating the ReplicateGlobalsPerAffinity pass from the Stream dialect to Flow, allowing replication logic to be managed within Flow and reducing cross-dialect dependencies. The feature was integrated into the Flow pipeline immediately after executable cleanup, gated by a rollout flag to support gradual deployment and risk-managed validation. The work, implemented in C++ using MLIR and compiler design principles, improved pipeline performance and specialization by localizing replication logic, enabling more targeted optimizations and streamlined architectural progression.