The research goal: can an assistant that lives inside a web page feel alive and useful without calling an AI model every few seconds? Everyone remembers the animated helper that popped up uninvited and understood nothing. The idea was not wrong. The technology did not exist, and the naive version today would cost a fortune in inference.
What we are trying to learn is how far deterministic code and pattern rules can carry the experience before a model is needed at all. The target is three to five AI calls per hour of active use, not three to five per minute.
Intelligence is layered in three tiers. Tier zero is pure code: the character's movement, its expression state machine, page scanning, and pathfinding between elements, running at zero cost. Tier one is a rules engine: a dozen recognized page patterns such as login, search, and settings, each with scripted guidance. Tier two is a small local model through Ollama, rate limited and cached, used only when the rules cannot answer or the user asks a direct question.
It ships as a Chrome extension that injects into any page, reads the page's structure, highlights the element it is talking about, and physically moves to it. Built standalone first, designed to plug into the production suite later.
What the lab built
At a glance
The extension loads on any page
Reads the page structure
Rules recognize the page type
Local AI, only when the rules cannot answer
Moves to the element and shows the way
Elsewhere in the lab
