Skip to content
Sentinel AI Sentinel AI

Sentinel Mesh

Sentinel Mesh is the gateway behind the API at sentinel.junaid.pk/v1. Every request flows through one governed pipeline.

Sovereign reasoning

The in-country model (qwen36-35b-a3b) is a reasoning model — it thinks before it answers. Sentinel captures that reasoning trace and surfaces it alongside the reply (a collapsible 💭 Thinking view in the Playground and Sandbox), so you get deep, deliberate answers and see the work — the reasoning stays in-country, on a sovereign model.

The pipeline

The Sentinel Mesh request pipeline A request passes through five stages in order: virtual-key authentication; quota and credit reserve; guardrails; sovereign routing to Sky47 Ascend inference inside Pakistan; and metering and settlement in PKR. IN-COUNTRY · pk-isb-1 1 Auth sk-mesh-… · sha256 scoped, RPM-capped 2 Quota & credit reserve against prepaid balance 3 Guardrails block · redact before dispatch 4 Sovereign route Sky47 Ascend in-country 5 Meter & settle priced in PKR hold settles REQUEST RESPONSE + USAGE
Every request takes the same path. Stage 4 is the only one that leaves the gateway: sovereign by default, and pinned in-country for anything marked sensitive, whatever the routing policy. External providers are opt-in per platform (see routing below) and absent entirely from the Qila build.
  1. Virtual-key auth — sk-mesh-… keys, stored as sha256, scoped (chat/analytics/admin), each with a hard per-key RPM cap.
  2. Quotas & credits — a per-tenant monthly token quota, plus a credit reserve against the prepaid balance (reserve/settle, so concurrent requests stay within the balance).
  3. Guardrails — blocked-terms and PII redaction before dispatch, with per-tenant policy overrides that only ever tighten the global floor, and a live events feed.
  4. Sovereign routing — the request runs on Sky47 Ascend in-country (qwen36-35b-a3b), always on the in-country model on the sovereign path.
  5. Meter & settle — the call is priced (wholesale × tenant markup × FX) and metered in PKR; the hold settles at retail.

Three protocols, one router

Three wire formats, one in-country pipeline Sentinel accepts requests in the OpenAI, Anthropic and Gemini wire formats, so existing SDKs work unchanged. These are request shapes the gateway serves, not external providers it calls: all three feed the same governed pipeline and, by default, the same in-country model. THE SHAPE YOU SEND — YOUR SDK, UNCHANGED OpenAI wire format POST /v1/chat/completions Anthropic wire format POST /v1/messages Gemini wire format POST /v1beta One pipeline auth · quota · guardrails route · meter · settle STREAMING ON /v1 In-country model qwen36-35b-a3b Sky47 Ascend · pk-isb-1 DEFAULT PATH These are request shapes Sentinel serves — not calls to those companies. Your request reaches the in-country model; routing anywhere else is opt-in and off by default.
Three front doors, one governed path. Point an existing OpenAI, Anthropic or Gemini SDK at Sentinel and it works unchanged — the wire format differs, everything behind it does not. These are formats Sentinel serves in-country; they are not the external providers described under routing below.

The Mesh speaks OpenAI (/v1/chat/completions), Anthropic (/v1/messages), and Gemini (/v1beta) wire formats over the same pipeline, with streaming on the OpenAI route. Any of those SDKs works unchanged.

Multi-provider routing — sovereign-first

One key, many models. The router ranks a pool of candidates and dispatches to the best fit, then falls back down the list on failure — the OpenRouter idea, rebuilt on sovereign ground:

  • Sovereign is the default, not the fallback — ask for nothing and you get the in-country model. A routing policy is something you opt into per request, so external routing can be enabled platform-wide without quietly moving anybody’s default traffic offshore.
  • Policies — cheapest (lowest wholesale per-token), fastest (rolling p50 latency the gateway measures itself), or use_case (the prompt is classified — code, analysis, chat — and matched to a model that fits).
  • Sovereign-first — the in-country model always leads, and any request marked sensitive is pinned in-country regardless of policy. External providers are opt-in per platform; the sovereign path stays in-country.
  • A provider registry, not a redeploy — operators add, key, enable, or disable providers and models from the admin console (/app/admin/engine); routing picks them up live. Any OpenAI-compatible endpoint works without an adapter — OpenAI and DeepSeek are onboarded that way. Keys are write-only: the console can replace one but never read it back.
  • Compiled out for defence — the Qila build ships with external routing removed at build time; the off-country code path is absent by construction.

Every reply reports the model that served it and whether it stayed sovereign, so routing stays transparent.

Billing you can resell

Every request is metered and priced. Operators get per-tenant, per-period invoices with wholesale, retail, and margin — so AI becomes a billed product with visible margin.

Capacity

Priority is defence › paid › free, with high availability across sovereign endpoints. Every key is capped — the Mesh is the sole rate-limiting authority.

See the Quickstart to make a call.