Software Engineer Begins 32-Week Journey From AI APIs to Local LLM Systems on NVIDIA DGX Spark

A software engineer and technical program manager has launched a 32-week self-directed program called 'From API to GPU,' aimed at transitioning from consuming AI via APIs to architecting local large language model systems. The project uses an NVIDIA DGX Spark as the primary compute machine, with a MacBook Pro serving as the control device for writing and remote access. In the first week, the author focused on understanding the DGX Spark's hardware environment, including its 20-core ARM64 CPU, GPU specifications, memory architecture, and CUDA setup, all verified through shell commands. A key early finding was that the MacBook and the DGX Spark run on different CPU architectures — x86_64 and ARM64 respectively — which has practical implications for software compatibility throughout the project. The full program is structured across eight phases, covering local model execution, ML fundamentals, transformer architecture, quantization, inference engineering, production deployment, RAG integration, and fine-tuning, with a parallel CUDA track beginning around week five.