Skip to content
Control Plane
How this demo works

Real routing. Simulated models.

An AI operations control plane decides whether each request runs on a local model or a cloud model, and enforces privacy, cost, quality and review rules around it. On this public site, every decision is real; only the model calls are simulated.

Compare modelsStart with a scenario

What runs for real

Every request takes the same path. One step is simulated here.

  1. 01RealClassifySensitivity scanner and task profile
  2. 02RealPolicyEvery model checked, fail-closed
  3. 03RealScore and routePrivacy, cost, task fit, reasoning, latency
  4. 04RealBudget GuardWorst case reserved before any call
  5. 05SimulatedModel executionRecorded replays where one exists
  6. 06RealEvaluationOn recorded text; simulated text is skipped
  7. 07RealReview and traceApproval queue and full request trace

Real on this site

  • Task classification and the sensitivity scanner (personal data, credentials, markings)
  • Governance policy evaluation for every model, including fail-closed blocking
  • Routing scores across privacy, cost, task fit, reasoning and latency
  • Budget Guard, worst-case cost reservation and fallback chains
  • Human-review rules, the approval queue and the request trace

Simulated on this site

  • Model inference: no model is called from this site
  • Local outputs: replays of real Gemma 4 12B and Qwen 3.5 9B runs recorded on the developer's Mac where one exists for the exact task; otherwise an explicit notice
  • Cloud outputs: illustrative text, always labeled simulated, never evaluated or recommended
  • Latency and cost of cloud models: the router's estimates at list price

A local install also

  • Discovers installed Ollama models and runs them on the device, at $0 in API fees
  • Optionally calls Claude through the Anthropic API, behind two explicit switches and a worst-case budget reservation
  • Measures real latency and throughput and feeds them back into routing
  • Keeps history, approvals and the spend ledger in SQLite

Why the public site can't

It deliberately cannot reach the developer's local Ollama runtime: there is no tunnel, proxy or exposed port to any private machine.

No provider credentials are used in Portfolio Mode. Paid API execution is disabled server-side, so no visitor can generate paid usage. Requests that ask for live execution are rejected with a 403, not quietly downgraded.

Anything you type is processed in memory by deterministic code and returned only to you; it is never added to the shared demo history.