Inside the engine.

The gate.

A message comes in. Before anything reaches the cloud, it hits a local machine. The gate asks 22 yes-or-no questions — is this a greeting? Does it contain a name? Is it asking for something complex? Fourteen get answered by pattern matching in under 5 milliseconds. No AI involved. The other eight need language understanding, so those get deferred to the first cloud call.

A greeting never leaves the building. A real question has already been classified, scored for complexity, and routed — before the cloud has seen a word.

Privacy.

While the gate classifies, a scanner looks for anything personal. Names, phone numbers, emails, addresses. When it finds them, it doesn't redact — it replaces. Real names become realistic pseudonyms. Real phone numbers become different real-looking numbers. The cloud gets a message that reads like natural language, just with someone else's identity. After the cloud responds, the local machine swaps everything back. The real data never left.

Memory.

Most AI remembers by keeping a log of everything and hoping the model finds the relevant part. This system doesn't do that. Every fact gets ten numerical scores — how important, how recent, what domain, how sensitive, how confident the source. A question triggers math against those scores. Only what's relevant gets pulled. Nothing more.

Four paths.

Not every question needs the same amount of work. The system decides which path based on the gate scores, not on how long the message is.

The modules.

Each piece does one thing. Green border means local execution — no API cost. Amber means an external model call.