Business Problem & Solution

Ten apps in this ecosystem run metered model stages, each with a prompt that someone tuned at some point. Without a control room, those prompts get edited blind in code, nobody knows which version is live, and a 'quick fix' to one app's prompt ships with no measurement. That is how an AI product quietly regresses.

ModelOps & PromptOps Studio is the control room. It shows which models are available in each mode, which prompt version is live per app, how each version scores on a fixture harness, and how recent runs behaved — so prompt changes are evaluated and promoted, not edited and hoped.

It proves the discipline: prompts are versioned with bodies and eval scores, an active version is explicit, promotion is an owner action that audits, and an eval can be run on demand against a fixture harness. The model catalog is honestly dual-mode — a Cloud model alongside OSS actives, a candidate, and a retired entry.

It simplifies the harness. Eval scores come from a deterministic fixture rather than a live benchmark run, and the eval call is a simulated metered round-trip. The studio is AI-assist over a real registry, not itself a model product.

Synthetic prompts only — five templates standing in for the ecosystem's metered apps (#23 BI, #05 review, #02 concierge, #06 content, #33 KYC), with invented bodies and scores. No production prompts, no proprietary model details, no PII. Figures are representative and labelled.