Platform
Cloud-agnostic. AWS, GCP, Azure, or bare metal. Operation is the client's responsibility.
The studio’s platform. Cloud-agnostic. Runs on AWS, GCP, Azure, or bare metal. Deployment is the client’s choice. Operation is the client’s responsibility.
§ 01 — Deploy targets
| Target | Form |
|---|---|
| AWS | EKS · GPU-class |
| GCP | GKE · Autopilot exempt |
| Azure | AKS · constrained subnet |
| Bare metal | k3s · air-gapped |
§ 02 — Components
Every component is replaceable. The studio supplies opinions; the client supplies decisions. Substitution is documented at handover.
| Component | Role | Substitute |
|---|---|---|
| Inference plane | Model serving · single tenant | vLLM / TGI / TensorRT-LLM |
| Trace bus | Append-only, signed | Kafka / NATS |
| Registry | Models, prompts, evals | OCI artifacts |
| Gateway | AuthN, rate, budget | Envoy / custom |
| Observability | Metrics, traces, logs | OTel · client choice |
§ 03 — Topology
Client VPC → Inference plane (vLLM · single tenant) → Model registry (read-only · signed)
§ 04 — Invocation
# Specify before you select.
$ aipe spec init my-system
$ aipe spec add boundary --vpc=client-prod --no-egress
$ aipe spec add trace --format=jsonl --signer=hsm
$ aipe spec add inference --target=eks --concurrency=64
# Verify, then build.
$ aipe spec verify
✓ boundary — isolated
✓ trace — signed, append-only
✓ inference — single-tenant, vLLM
✓ handover document — generated
$ aipe build && aipe handover