Andrija Bosnjakovic developed a core feature for the llvm/torch-mlir repository, focusing on the integration and lowering of stochastic ReLU variants within the Torch MLIR dialect. He implemented support for torch.aten.rrelu_with_noise and its backward operation, enabling more robust and optimized machine learning workflows. Using C++ and Python, Andrija refined the TorchToLinalg lowering path, directly addressing correctness and reliability issues in downstream pipelines. His work enhanced the end-to-end readiness of MLIR-based tooling for models utilizing stochastic ReLU, facilitating easier adoption and improved performance. The depth of his contribution reflects strong expertise in C++ development, MLIR, and machine learning.