AJRouter Engine: Run and manage llama.cpp and audio.cpp on Windows

Hi, I’m AJ. AJRouter grew out of my Windows local AI setup becoming too many manual steps: starting services, switching GGUF profiles, watching VRAM, and keeping text and audio workloads from colliding. I built a control plane around llama.cpp and audio.cpp so the setup is repeatable and coding agents can use a stable local API. On my RTX 5060 Ti 16 GB, one surprise was that pushing MTP draft depth higher made a five-task test slower. I documented the build, measurements and their limits here: https://ajthe.dev/case-studies/p... This is a paid Windows download with the runtimes included; you supply your own GGUF models. If you run local models, what part of operating the stack costs you the most time?