Before we route anyone else’s AI spend, we ran the architecture on our own operations. This is the teardown: what it is, what it runs, and why the cost-routing holds up.
The fleet is hybrid by design. Open-weight models running on owned hardware handle the bulk of the day-to-day work. Cloud frontier models are never the default — they’re routed in through a broker only for the tasks that justify their cost. A model-tiering policy decides which model each job gets, guard checks audit what the agents do, and a memory layer keeps context across sessions.
Routing work to cheaper models only saves money if quality holds. The part that makes it safe isn’t the router — it’s the evaluation gate built from your real tasks, so you can see whether a smaller model actually did the job before you trust it with the job.
That’s the whole discipline: measure the traffic, classify it, prove the cheaper path on your own work, then route. Escalate to a frontier model only where the eval says you need it. It’s the same cost-routing architecture the industry began marketing heavily in 2026 — the difference is that it has been running here in production, not sitting in a slide.
A founder who runs the architecture he sells is the whole point.