Multi-model routing with cost governance
One model for every task overpays. The router resolves each task class to a model by data, not a hardcoded choice, and a deterministic governor proposes cheaper routes as usage patterns settle.
How it works
Every LLM call site resolves its model through a task class — an interview question and a full spec synthesis are allowed to route differently by design.
Each route carries a fallback chain, so a provider outage degrades a task rather than failing it outright.
Every call's usage and estimated cost rolls up per day, task class, and model — the data a routing decision should be based on.
Bulk work runs through bounded-concurrency workers with cache-friendly prompts and provider batch mode where it's offered — and the batch discount shows up in your spend view.
A deterministic sweep over usage rollups proposes route downgrades with estimated savings — a human applies the change by writing the route.
The agents behind it
Human control
A human applies any route change directly in the platform settings; the governor only ever proposes.