Managed execution
Sovereign AI Machine
The execution component of the platform: it takes a request from the control plane, routes it to Sovereign AI, records metadata and returns the answer. Offered to customers as Tier B — managed execution — once the cloud copy is adapted.
What exists today
Gateway
BuiltA FastAPI gateway with /health, /v1/execution-route, POST /v1/respond and POST /v1/respond/stream; loopback only; external image suppliers off in the cloud copy.
Tests and acceptance
Built151 unit tests and 14 named acceptance checks (I1 … R3) on the Machine the cloud copy is forked from.
Routes to Sovereign AI
BuiltThe cloud copy routes every turn to Sovereign AI: the inference provider of the platform, nothing else.
No external supplier
BuiltThe cloud copy carries no route to any external AI supplier; the fallback that other systems call a feature does not exist here, by contract.
Status labels — MEASURED — exists and was measured on the stated date by the stated method. BUILT — exists and works, not yet measured under load. PLANNED — designed, not yet built. A line without a label is not a claim of this platform.
Two tiers
Tier A — direct API
The default: your client calls the OpenAI-compatible API of the control plane, which authenticates, meters and forwards the request to Sovereign AI. Streaming, tools and structured output as on the API page.
Tier B — managed execution
BuiltFor routed, multi-step work: the control plane hands the request to the Sovereign AI Machine copy, which executes it against Sovereign AI and returns the result with metadata. Planned for the second phase.
Managed execution — the contract
What Tier B promises is written in the contract between the control plane and the Machine copy. Each line carries its status.
Tenant headers
BuiltOn every call the control plane sets the organisation, project and key-prefix identifiers and a request id; the Machine never accepts them from a public caller.
Honest error states
BuiltBACKEND_BUSY when the GPU is held by a media job, UPSTREAM_DOWN when Sovereign AI is down — mapped by the control plane to the public error contract with a Retry-After.
Streaming through the Machine
BuiltManaged execution streams: the copy relays the inference gateway's chunks and the final usage, and decides the status before the first byte, so a busy GPU is an honest 503. Tier A streams directly from the inference gateway.
Tier B — managed execution
BuiltRouted, multi-step work for a customer tenant: the control plane hands the request to the Machine copy, which executes it against Sovereign AI and returns the result with metadata.
Isolation
Private network only
BuiltLoopback in development, a private network in the cloud; the Machine is never reachable from the public internet.
Identity set only by the control plane
BuiltA public caller cannot name a tenant: the identifiers are set by the control plane after authentication and ignored from anywhere else.
Audit without content
BuiltThe Machine writes the three identifiers and the request id into its audit line and never the request content.
The Original Machine is never exposed
BuiltMedia Holding's own internal Machine is a separate system. Nothing on this platform gives a customer access to Media Holding's internal tools, routes or data.
What customers never get
Access to Media Holding's internal systems. Routes to external AI suppliers. The image generation and the scheduling of the internal Machine. The public API exposes a limited customer surface and nothing of the Original Machine — by directive and by construction.