Demo: Smart Home Assistant

A Noul detects whether a request asks for more than one action; if so, an LLM splits it into atomic commands that TypeSafe evaluates individually. Otherwise a request resolves through four questions: category, target domain, device type, and action.

The page is an architecture write-up plus a Loom video with no code shown. The demo is a Vite/React single-page app, and the source was promised for GitHub at release rather than linked.

Each request is evaluated against a long list of questions, most of which turn out irrelevant. For turn off all of the lights in the house, the action question is written on the assumption the user is commanding lights and is asked before that is known, so everything evaluates in parallel and code filters the irrelevant answers afterwards.

Sequential calls minimize the number of questions but are much slower and more expensive than batching them into one upfront call. You would ask the category question, wait, then the domain question, wait again, then the device question, and only then the action question.

It handles the two steps that need generated strings: splitting a compound request into atomic commands, and answering general-information or conversational queries with a freeform response. The TypeSafe pass is so much faster than the LLM response that it adds negligible latency to the turn.

Because requests with known deterministic behavior resolve in one batched TypeSafe call that is far faster and cheaper than sequential LLM steps. The LLM is kept for the two steps that need generated strings: splitting a compound request into atomic commands and answering general-information or conversational queries.

A tool call with its arguments filled from plain language: the four questions resolve the request category, target domain, device type and the action to take, such as smarthome command, whole house, lights and turn off. Requests that are not commands fall back to a freeform LLM answer.