Most of what AdaDo does runs on hosted models we license and call over an API — Anthropic's Claude, Qwen through DeepInfra. This page is about the one thing that's genuinely ours: trained in-house, running our own weights, for our own pipeline.
A small, genuinely trained routing classifier: TF-IDF feature extraction feeding a one-vs-rest softmax logistic regression, trained from scratch via gradient descent (numpy, no third-party ML library) on our own distillation dataset. Its job is narrow and specific — given an incoming request, decide which of five actions AdaDo's engine should take: route to a skill, handle it ad hoc, let it run free, proceed without intervention, or pull from knowledge. That's the whole scope.
It isn't a chat model, isn't instruction-tuned, and doesn't generate text in any general sense. It's a classifier that was trained to do one job inside our own pipeline, and it does that job on commodity CPU instead of burning a call to a hosted LLM for a decision that doesn't need one.
The honest shape of it: we build the system around the models, not the models themselves — except for the one narrow piece where training something small ourselves was the right call instead of paying an API for a five-way classification.