TensorFlow GPU Inference Refactoring & Build

I have an existing GPU-accelerated network component that still compiles, but its inference logic is tangled and hard to maintain. I want the codebase cleaned up and modernised while keeping the same outputs and speed. Here is what I need from you: • Compile the current project on a recent CUDA-enabled Linux environment so we start from a reproducible build. • Re-architect the inference path, shifting it to clear, modular functions that follow TensorFlow best practices. The focus is code refactoring, not new features. • Preserve existing model weights and performance; any change that alters numerical results must be explained. • Deliver a git repo containing the refactored source, build scripts, and a short README that documents dependencies, compile flags, and how to run a quick sanity test. If you are comfortable navigating CUDA projects but prefer writing clean TensorFlow code, this should be straightforward. Let me know your typical turnaround time and a brief outline of how you would approach refactoring the inference logic.