Skip to content
Sentinel AI Sentinel AI

Sovereignty

Sovereignty is a build guarantee, not a setting.

Sovereign boundary: in-country inference, no external egress Client requests enter the Sentinel Mesh gateway. Sensitive and defence traffic is pinned to Sky47 Ascend inference inside Pakistan, in region pk-isb-1. External model providers sit outside the boundary, off by default; on the sovereign path no request reaches them, and in the Qila defence build that path is removed from the binary entirely. PAKISTAN · pk-isb-1 Clients SDK · app · console Sentinel Mesh auth · guardrails meter · route /v1 · /v1beta Sky47 Ascend qwen36-35b-a3b 35B · 128K ctx 2-card NPU slice not export-controlled External model providers OFF BY DEFAULT · NEVER SENSITIVE ON THE SOVEREIGN PATH Sensitive and defence traffic is pinned in-country, whatever the routing policy. In the Qila defence build the external path is compiled out — the code does not exist. Serving the OpenAI, Anthropic and Gemini wire formats is not calling those providers.
The dashed line is the whole product. Requests reach the in-country model and stop there. External model providers sit outside the boundary, off by default and never serving sensitive traffic — and in the Qila build that outward path is not in the binary at all. Note this is about where a request is routed, not which API format you call: Sentinel serves the OpenAI, Anthropic and Gemini wire formats in-country (see Quickstart).

In-country inference

Inference runs on Sky47 Ascend NPUs inside Pakistan (region pk-isb-1). Requests to the sovereign path never leave the country. The live model is qwen36-35b-a3b — a 35B model with a 128K context window, serving on a 2-card Ascend slice.

Zero external egress

Requests default to the in-country model — routing to anything else is opt-in per request, so enabling the external pool never silently moves existing traffic. Sensitive and defence traffic routes only to the in-country endpoint whatever the policy. The Sentinel Qila defence build is air-gapped: external-provider routing is compiled out of the binary, not toggled off — the code path does not exist.

The dependency nobody prices in

Frontier model capability is supplied by a small number of foreign providers, and access to it is not a technical fact — it is a commercial and political arrangement. Terms, pricing, rate limits, model availability and the export policy underneath them are all set somewhere else, and can change without notice and without appeal.

That is a switch, and someone else owns it. Most AI strategies assume it will never be thrown, because it usually isn’t — right up until the moment a policy, a sanction or a licence review makes the decision for you. A capability that can be withdrawn is not a capability you have; it is one you are renting under terms you do not control.

Sovereignty is the answer to that, not to a privacy checkbox. The point of running inference in-country, on hardware that is not export-controlled, is that no one outside the country can decide whether you still have it tomorrow.

Ascend, not export-controlled silicon

Sentinel runs on Huawei Ascend hardware, independent of the US chip restrictions that constrain regional competitors. The model, the weights and the machines are inside the country and inside the operator’s reach — which is what makes the guarantee above structural rather than contractual.

Capacity and control

  • Priority: defence › paid › free.
  • Every key is capped: per-key RPM limits and daily caps; no uncapped keys.
  • Metered: each call is credit-metered in PKR, with guardrail checks on the sovereign path.

Governance

Sentinel Watch provides a metadata-only audit feed and a live sovereignty posture, with NPU and throughput dashboards in Grafana.