Dragan Mladjenovic engineered robust GPU backend and build system enhancements across the ROCm/xla and tensorflow/tensorflow repositories, focusing on performance, compatibility, and maintainability. He delivered dynamic build configuration, atomic operation optimizations, and convolution performance improvements by leveraging C++, LLVM, and ROCm integration. Dragan addressed cross-version compatibility by implementing dynamic SONAME detection and upgraded bitcode libraries to support new graphics architectures. His work included thread-safety improvements for LLVM command line handling and in-process LLD linking to reduce build overhead. Through careful refactoring and autotuning backend development, Dragan improved runtime stability and streamlined CI, demonstrating depth in low-level optimization and compiler development.