Operating model

How Foundation is delivered, where platform and tenant workloads run, and who changes what.

Agentic Friendly

Foundation is operated as one shared platform layer. Platform engineers change it through rendered configuration and GitOps, platform administrators change tenants through the Portal and the BSQAI API, and AI and Data services consume it without operating their own copies of its services.

Core principle

Foundation owns platform concerns. Higher layers own product and use-case concerns.

  • Foundation focuses on secure execution, exposure, persistence, observability, and delivery.
  • AI focuses on AI platform services and developer-facing capabilities.
  • Data focuses on enterprise data systems and advanced data operations.

How the platform is delivered

The platform CLI renders the desired state of every component from platform.toml, the components' variables.toml files, and environment overrides, then pushes it to the manifests repository. Argo CD reconciles it into the cluster through four parent applications:

  • common: networking, identity, secrets, storage, databases, observability, and the tenant operator
  • coreai: Portal, BSQAI API, MLflow, Milvus, Temporal, and MCP management
  • inference: KServe, KAI Scheduler, the model gateways, and runtime profiles
  • proai: Data and ML services, Argo Workflows, Argo Events, and Kyverno

A few releases are bootstrap prerequisites and stay outside the GitOps repository. The platform CLI installs them directly with bsqai helm-apply:

ReleaseWhen it is installed
argocdAlways. Argo CD must exist before it can reconcile the other applications.
gpu-operatorOn GPU clusters that use the platform's NVIDIA integration
calico, then calico-installationOn clusters that need the platform to provide a network policy engine

Bootstrap releases change only when bsqai helm-apply runs again. See GitOps workflow for the render and reconcile flow, sync ordering, and Argo CD behavior.

Where workloads run

WorkloadsNamespaceManaged by
Platform services, one namespace per component, such as keycloak, llm-backend, or grafanaPlatform namespacesArgo CD, from the rendered manifests
Models of the Global OrganizationkserveThe BSQAI API model installer
A tenant's models, Managed MCP Servers, and private MLflowtenant-<slug>, one per tenantThe BSQAI Tenant Operator prepares the namespace; the BSQAI API places models and MCP servers in it
Agent sandbox workspaces, where the site enables the sandboxcoreai-sandboxesThe sandbox controller, on behalf of the BSQAI API

The tenant operator gives each tenant namespace a ResourceQuota, a LimitRange with container defaults, and a default-deny network policy. Tenant users have no Kubernetes access to these namespaces: they work through the Portal and the BSQAI API.

Shared core services

Foundation runs a few services that every product area reuses instead of building its own:

  • Temporal runs durable workflows, such as tenant provisioning and deletion, file ingestion, and deep research.
  • Argo Workflows and Argo Events run multi-step pipelines and event-driven automation.
  • Kyverno and Kubernetes ValidatingAdmissionPolicy resources enforce admission rules, such as the serving runtimes, resource profiles, and GPU queues a tenant model may use.
  • Argo CD keeps every platform component aligned with the manifests repository.

Who changes what

ChangeMade byThrough
Platform configuration, enabled components, and versionsPlatform engineerplatform.toml and overrides, bsqai apply, bsqai push-manifests, then Argo CD
Bootstrap releasesPlatform engineerbsqai helm-apply
Tenant creation and deletion, tenant services, deployment policy, quotas, and BCU allowancePlatform administratorAdministration in the Portal, or the BSQAI API, such as /v1/admin/tenants; Temporal and the tenant operator apply tenant changes
Membership, roles, federation, model access rules, caps, and tenant settingsTenant administratorAdministration in the Portal, or the BSQAI API

See Tenant lifecycle control plane for the tenant routes and Responsibility boundaries for the full role list.

How other layers consume Foundation

  • An AI service uses Foundation ingress, identity, storage, inference, and telemetry.
  • A workflow-oriented service uses Foundation workflow engines and storage.
  • A user-facing application uses the same DNS, TLS, and access patterns as the rest of the platform.

On this page