Most people interact with LLMs through an API and never think about what sits between the model weights and the HTTP response. A look at the two-phase execution model, the KV cache, and why llama.cpp and vLLM solve fundamentally different problems.

How writing the spec first — and letting it govern the entire development cycle — maps naturally to Scrum roles, keeps AI-assisted code generation honest, and delivers real velocity to the SDLC.