Feature Architecture

AI-Platform Architecture		status: draft security: NO safety: ASIL_B
version: 1 realizes: wp__feature_arch
doc__ai_platform_architecture		Generic Document

AI-Platform		status: valid security: NO safety: ASIL_B
version: 1 satisfies: feat_req__ai_platform__agent_execution, feat_req__ai_platform__embedded_efficiency, feat_req__ai_platform__model_lifecycle, feat_req__ai_platform__model_verification, feat_req__ai_platform__observability, feat_req__ai_platform__qnx_linux_support, feat_req__ai_platform__safety_backends, feat_req__ai_platform__static_backend, feat_req__ai_platform__workloads_execution includes: logic_arc_int__ai_platform__inference_if
feat__ai_platform		Feature

Interfaces

Inference

The inference interface is the application-facing entry point of the AI Platform. It is intentionally narrow: it exposes the common denominator of the supported backends and omits vendor-specific options, accelerator selection and memory placement, which remain platform-internal.

Inference Interface		status: valid security: NO safety: ASIL_B
version: 1 fulfils: feat_req__ai_platform__safety_backends, feat_req__ai_platform__static_backend, feat_req__ai_platform__workloads_execution included by: feat__ai_platform includes: logic_arc_int_op__ai_platform__cancel, logic_arc_int_op__ai_platform__describe, logic_arc_int_op__ai_platform__invoke, logic_arc_int_op__ai_platform__invoke_async, logic_arc_int_op__ai_platform__load, logic_arc_int_op__ai_platform__unload
logic_arc_int__ai_platform__inference_if		Logical Interface & Feature Interface View

load		status: valid security: NO safety: ASIL_B
version: 1 included by: logic_arc_int__ai_platform__inference_if
Resolves a model by identity and version constraint and returns a model handle. Input and output descriptors are fixed once the model is loaded.
logic_arc_int_op__ai_platform__load		Logical Interface Operation

describe		status: valid security: NO safety: ASIL_B
version: 1 included by: logic_arc_int__ai_platform__inference_if
Returns the input and output descriptors of a loaded model, consisting of name, data type, shape and layout.
logic_arc_int_op__ai_platform__describe		Logical Interface Operation

invoke		status: valid security: NO safety: ASIL_B
version: 1 included by: logic_arc_int__ai_platform__inference_if
Executes a loaded model synchronously within a given deadline. Exceeding the deadline is reported as an error instead of blocking the caller indefinitely.
logic_arc_int_op__ai_platform__invoke		Logical Interface Operation

invoke_async		status: valid security: NO safety: ASIL_B
version: 1 included by: logic_arc_int__ai_platform__inference_if
Starts an execution and returns a handle that allows waiting for completion.
logic_arc_int_op__ai_platform__invoke_async		Logical Interface Operation

cancel		status: valid security: NO safety: ASIL_B
version: 1 included by: logic_arc_int__ai_platform__inference_if
Requests termination of a pending asynchronous execution and releases its resources.
logic_arc_int_op__ai_platform__cancel		Logical Interface Operation

unload		status: valid security: NO safety: ASIL_B
version: 1 included by: logic_arc_int__ai_platform__inference_if
Releases a model handle and the resources associated with it.
logic_arc_int_op__ai_platform__unload		Logical Interface Operation

Interface Sketch

The following sketch illustrates the interface from an application perspective. It is illustrative and does not fix the final signatures.

namespace score::ai {

// Describes one input or output of a model. Shapes are fixed after Load.
struct TensorDesc {
  std::string_view name;
  DataType type;             // e.g. Float32, Float16, Int8, UInt8
  std::span<const int64_t> shape;
  Layout layout;             // e.g. NCHW, NHWC, Linear
};

// Opaque handle to a platform-managed buffer. Allows zero-copy for data that
// already resides in accelerator memory (e.g. camera output).
class Buffer;

class Model {
 public:
  std::span<const TensorDesc> Inputs() const;
  std::span<const TensorDesc> Outputs() const;

  Result<void> Invoke(std::span<const Buffer> inputs,
                      std::span<Buffer> outputs,
                      Deadline deadline);

  Result<Execution> InvokeAsync(std::span<const Buffer> inputs,
                                std::span<Buffer> outputs,
                                Deadline deadline);
};

// Resolves a model by identity and version constraint; the platform selects the
// backend and the accelerator.
Result<Model> Load(ModelId id, VersionConstraint constraint);

}  // namespace score::ai

Properties implied by this interface:

• Identity instead of paths: applications reference a model by identity and version constraint, so model artifacts can be updated, signed and rolled back without changing the application.

• Fixed shapes after load: descriptors are known after loading, which allows buffer allocation and timing analysis before the first execution.

• Explicit deadlines: every execution carries a deadline, so a missed deadline is a defined, reportable error.

• Buffer handles instead of pointers: data may stay in accelerator memory, which keeps zero-copy paths possible without exposing the memory model to the application.

• No backend in the signature: the selected runtime (e.g. ONNX Runtime, TensorRT, QNN) is a deployment decision and does not appear in application code.