Three tiers, cheapest capable one wins
Every request is classified before it runs. Most never leave the property. The tiers degrade gracefully in both directions: if a local tier is unreachable, the request moves up rather than failing.
On-box classification
A small model that decides where each request should go, plus semantic search and cheap templating. Milliseconds, no accelerator required.
Local inference on Apple Silicon
Large open-weight models served natively over unified memory. Conversation, drafting, summarizing, first-pass review. This tier does the bulk of the work.
Frontier escalation
Reserved for genuine judgment: production changes, security, anything irreversible. Certain categories are hard-routed here and never left to the classifier.
How it’s built
Local by default
Escalating off-box is a deliberate decision the system has to justify, not the default path.
Fails forward, never hangs
Every tier has a bounded timeout and a next step. An unreachable model degrades the route; it doesn’t drop the task.
Guardrails aren’t prompts
High-stakes categories are enforced at the process boundary, not requested politely in a system prompt.
Measured, not assumed
Routing decisions are logged and reviewed. If a tier isn’t earning its place, the data says so.
This is private infrastructure
Aria is personal infrastructure, not a product. There is no sign-up, no waitlist, and no public endpoint, this page is a description of the system, not a way in.