Cardix

Architecture

Inside your perimeter

Cardix runs as a single service within your network. Applications connect via an OpenAI-compatible endpoint; Cardix handles classification, policy, selection, and dispatch.

YOUR PERIMETER — VPC / on-prem / air-gappedYour applicationsunchanged OpenAI-compatible SDKCardixclassify · policy · select · dispatchRoutingLearningObserveonly approved outbound trafficOn-prem modelsvLLM · Ollama · private deploysApproved cloud APIsOpenAI · Anthropic · Bedrock

Planes

Three layers, one purpose

Routing, learning, and observability work together — without exposing your data or blocking your responses.

Routing plane

Classify sensitivity and task, apply policy, select the best eligible model, and dispatch — synchronously on every request.

Learning plane

Improves routing over time from traffic signals. Runs asynchronously and never blocks the response.

Observability plane

Full audit trail, spend vs baseline comparison, and visibility into blocked demand.

Request journey

What happens on every request

Five steps, milliseconds — policy before routing, every time.

  1. 01

    Authenticate

    Validate the caller and attribute the request to your organization.

  2. 02

    Classify

    Determine data sensitivity and task category for the prompt.

  3. 03

    Policy

    Evaluate residency rules and produce allowed deployment locations. Fail closed if no rule matches.

  4. 04

    Select

    Choose the best model from your eligible catalog for this request.

  5. 05

    Dispatch

    Route to the chosen backend and return the response, streaming or buffered.