@qnsp/ai-sdk

TypeScript client for the QNSP AI Orchestrator. Provides workload submission, model deployment,
inference, training, and artifact management helpers with built-in tier enforcement.

Installation

``bash pnpm add @qnsp/ai-sdk`

`Authentication`

The client authenticates with a QNSP service token (PQC-signed JWT) issued by the auth service. Pass the token via thetoken option. You may optionally set tierif you are enforcing pricing entitlements on the client before making API calls.

`ts import { AiOrchestratorClient } from "@qnsp/ai-sdk";

const ai = new AiOrchestratorClient({ baseUrl: "https://ai.qnsp.cuilabs.io", token: process.env.QNSP_SERVICE_TOKEN!, tier: "dev-pro", });`

`Tier requirements`

| Capability | Minimum tier | Notes | |------------|--------------|-------| | Basic inference (invokeInference, deployModel) | dev-pro | Throws TierError if tieris lower | | Secure enclaves / GPU workloads |enterprise-standard| Triggered when workloads request GPU resources | | AI training / fine-tuning |enterprise-pro | Automatically enforced when workload name contains training or fine-tune |

The SDK relies on @qnsp/shared-kernel's checkTierAccesshelper, so upgrading the tier in the constructor immediately unlocks the relevant calls.

`Usage example`

`ts import { AiOrchestratorClient } from "@qnsp/ai-sdk";

const ai = new AiOrchestratorClient({ baseUrl: "https://ai.qnsp.cuilabs.io", token: process.env.QNSP_SERVICE_TOKEN!, tier: "enterprise-pro", });

// Register an artifact const artifact = await ai.registerArtifact({ tenantId: "tenant_123", modelName: "quantum-guardian", digest: "sha3-512:...", storageUrl: "qnsp://storage/documents/abc", });

// Submit an enclave-enabled training workload await ai.submitWorkload({ tenantId: "tenant_123", name: "guardian-training", resources: { cpu: 8, memory: "64Gi", gpu: 2 }, containerImage: "ghcr.io/qnsp/training:latest", command: ["python", "train.py"], artifacts: [{ artifactId: artifact.artifactId, mountPath: "/models", accessMode: "read" }], env: { DATASET_PATH: "/datasets/latest" }, });`

`Telemetry`

All requests emit OpenTelemetry spans when you wrap fetchImplwith your tracing/export logic. Future releases will exposecreateAiClientTelemetry, matching other SDKs.

`Related documentation`

- Developer onboarding guide - AI orchestrator service plan - Pricing & tier limits

`License`

Licensed under the Apache License, Version 2.0. See LICENSE`.

Authentication

The client authenticates with a QNSP service token (PQC-signed JWT) issued by the auth service.
Pass the token via the

token option. You may optionally set tier

 if you are enforcing pricing
entitlements on the client before making API calls.

ts
import { AiOrchestratorClient } from "@qnsp/ai-sdk";

const ai = new AiOrchestratorClient({
  baseUrl: "https://ai.qnsp.cuilabs.io",
  token: process.env.QNSP_SERVICE_TOKEN!,
  tier: "dev-pro",
});

Tier requirements

| Capability | Minimum tier | Notes |
|------------|--------------|-------|
| Basic inference (

invokeInference, deployModel) | dev-pro | Throws TierError if tier

 is lower |
| Secure enclaves / GPU workloads |

enterprise-standard

 | Triggered when workloads request GPU resources |
| AI training / fine-tuning |

enterprise-pro | Automatically enforced when workload name contains training or fine-tune |

The SDK relies on @qnsp/shared-kernel's checkTierAccess

 helper, so upgrading the tier in the
constructor immediately unlocks the relevant calls.

Usage example

ts
import { AiOrchestratorClient } from "@qnsp/ai-sdk";

const ai = new AiOrchestratorClient({
  baseUrl: "https://ai.qnsp.cuilabs.io",
  token: process.env.QNSP_SERVICE_TOKEN!,
  tier: "enterprise-pro",
});

// Register an artifact
const artifact = await ai.registerArtifact({
  tenantId: "tenant_123",
  modelName: "quantum-guardian",
  digest: "sha3-512:...",
  storageUrl: "qnsp://storage/documents/abc",
});

// Submit an enclave-enabled training workload
await ai.submitWorkload({
  tenantId: "tenant_123",
  name: "guardian-training",
  resources: { cpu: 8, memory: "64Gi", gpu: 2 },
  containerImage: "ghcr.io/qnsp/training:latest",
  command: ["python", "train.py"],
  artifacts: [{ artifactId: artifact.artifactId, mountPath: "/models", accessMode: "read" }],
  env: { DATASET_PATH: "/datasets/latest" },
});