Umesh Malesh built and enhanced core backend and compiler features for Tenstorrent’s tt-mlir and tt-xla repositories, focusing on machine learning workloads and JAX integration. He implemented the dot_general operation in TTIR, decomposing it into matmul, permute, and reshape to support broader linear algebra use cases and enable linear regression compilation. Umesh expanded test coverage for JAX numeric operations, introducing tunable PCC validation and MLP training tests to improve reliability. He also developed StableHLO to TTIR conversion for random number generation and fixed MLIR lowering shape handling, using Python, MLIR, and C++ to address correctness and stability issues.