Then I put the engine inside a live conversation, reasoning in real time while people were still talking.
By 2025 the frontier moved from answering after the fact to reasoning in the moment. A live meeting is an unforgiving test: the engine has to follow an unstructured conversation as it happens, track what matters, and surface it before the moment passes.
Reasoning over stored knowledge is one thing. Reasoning against a conversation while it is still happening is another, and much harder.
So I built a layer that sits on a live meeting and works in real time, following the conversation, tracking the open questions, the commitments, the things left unresolved, and pulling in the context a person would need but would never remember to look up mid-sentence. The engine reasoning at the speed of the room.
That is the version of the thesis I care about most. Not software that answers when you finally ask, but software that reasons alongside you and surfaces the decision before you even knew to make it. Reasoning before a human has to.