# Reference Architecture (/docs/runtime/reference-architecture)



The Runtime layer is not a single component. It is a platform architecture made of several cooperating domains.

Architecture View [#architecture-view]

<Mermaid
  chart="flowchart TB
    U[&#x22;Users and Client Systems&#x22;] --> DNS[&#x22;External DNS&#x22;]
    DNS --> IN[&#x22;Ingress and API Exposure&#x22;]
    IN --> APISIX[&#x22;APISIX&#x22;]
    IN --> NGINX[&#x22;NGINX&#x22;]
    IN --> LB[&#x22;MetalLB&#x22;]

    APISIX --> ID[&#x22;Identity and Access&#x22;]
    NGINX --> ID
    ID --> KC[&#x22;Keycloak&#x22;]
    ID --> FGA[&#x22;OpenFGA&#x22;]
    ID --> BAO[&#x22;OpenBao&#x22;]

    APISIX --> RUN[&#x22;Runtime Services&#x22;]
    NGINX --> RUN
    RUN --> INF[&#x22;Inference&#x22;]
    RUN --> WF[&#x22;Workflows&#x22;]
    RUN --> DATA[&#x22;Operational Data&#x22;]

    INF --> KAI[&#x22;KubeAI&#x22;]
    INF --> VLLM[&#x22;vLLM&#x22;]
    INF --> FW[&#x22;FasterWhisper&#x22;]

    WF --> AE[&#x22;Argo Events&#x22;]
    WF --> TMP[&#x22;Temporal&#x22;]

    DATA --> PG[&#x22;PostgreSQL&#x22;]
    DATA --> PGA[&#x22;PgAdmin&#x22;]
    DATA --> MINIO[&#x22;MinIO&#x22;]

    RUN --> OBS[&#x22;Observability&#x22;]
    OBS --> GRAF[&#x22;Grafana&#x22;]
    OBS --> PROM[&#x22;Prometheus&#x22;]
    OBS --> LOKI[&#x22;Loki&#x22;]
    OBS --> ALLOY[&#x22;Alloy&#x22;]
    OBS --> TEMPO[&#x22;Tempo&#x22;]

    GIT[&#x22;GitLab&#x22;] --> CD[&#x22;Argo CD&#x22;]
    REG[&#x22;Harbor&#x22;] --> CD
    CD --> RUN"
/>

Capability Domains [#capability-domains]

| Domain           | Main role                                         | Example components                       |
| ---------------- | ------------------------------------------------- | ---------------------------------------- |
| Network          | Expose, publish, and route traffic                | MetalLB, NGINX, APISIX, External DNS     |
| Security         | Authenticate, authorize, and protect secrets      | Keycloak, OpenFGA, OpenBao, cert-manager |
| Inference        | Run production model workloads                    | KubeAI, vLLM, FasterWhisper              |
| Workflow engines | Execute event-driven and durable processes        | Argo Events, Temporal                    |
| Data plane       | Store operational state and artifacts             | PostgreSQL, PgAdmin, MinIO               |
| Observability    | Collect and explore platform signals              | Grafana, Prometheus, Loki, Alloy, Tempo  |
| Delivery         | Move artifacts and desired state into the cluster | GitLab, Harbor, Argo CD                  |

Design Intent [#design-intent]

This architecture allows BullSequana AI to keep infrastructure concerns explicit and reusable. Higher layers can then consume Runtime services without rebuilding the same operational patterns for each product area.