AI inference for Adfinia, pinned to a local LLM you control. Customer data never leaves your environment. No outbound call to a US-hosted model API. Toggleable per tenant; one switch, full audit trail.
Generic marketing platforms route AI prompts to a US-hosted model. Your prompts include customer data. That's a lot of trust to outsource to someone else's transit network.
Prompts, completions, embeddings, and any retrieved context never leave the geographic boundary you select — including the AI hop. The model runs on your GPU, in your network.
No SaaS LLM provider sub-processor for sovereign-flagged tenants. Your DPA doesn't need to list a model vendor; your audit doesn't need to ask one for a SOC 2.
Built for procurement, defence, healthcare, and regulated public-sector tenants that can't accept outbound model-API traffic. We pin the AI router, then publish proof.
Every AI feature in Adfinia goes through one internal service — the AI Gateway. Sovereign mode is a flag on the tenant that switches that gateway's upstream from the cloud model API to a local inference service.
cloud-ai-provider.api → completion
local-llm.svc.cluster.local → completion
SELECT COUNT(*) FROM ai_invocations WHERE destination LIKE '%cloud-ai-provider%' AND tenant_id = '…' must return zero for a sovereign tenant.
Sovereign isn't a feature flag you flip in a meeting and flip back the next morning. It's a posture change with operational and contractual consequences — so we put a 30-day cooldown on both directions of the switch.
Once you flip on Sovereign, the next AI invocation routes to your local LLM. The flag itself is locked for 30 days — you cannot switch back to cloud sooner. We use the window to confirm your model is stable, your performance is baselined, and your audit trail shows zero cloud egress.
If you want to switch back to cloud (typically because you've outgrown your on-prem capacity), the flag is locked for 30 days the other way. The window gives your security team time to update DPAs, re-list sub-processors, and notify regulators where required.
The cooldown is enforced at the platform layer — not a soft setting. The admin console shows the unlock date; a super-admin can force-unlock with a typed reason that hits the audit log and a #admin-audit Slack notification.
Sovereign tenants pin both data residency and AI inference to one jurisdiction. The current options:
In-country inference for PDPL-regulated workloads and government tenants. Mistral, Falcon, and licensed Llama variants supported.
In-country inference for DPDP-regulated workloads. Aligns with MEITY guidance on cross-border data flow.
EEA-resident inference for GDPR-restricted workloads. Compatible with EU AI Act transparency obligations.
In-country inference for SDAIA-regulated workloads. Available on contract; provisioned via local partner.
In-country inference for ODPC-aligned workloads. Roadmap; talk to us if this is a procurement requirement.
Deploy the inference service inside your own Kubernetes cluster. We bring the model wrapper; you bring the GPU.
Switching jurisdictions inside Sovereign mode is itself subject to the 30-day cooldown. See the DPA for data residency and transfer details.
Sovereign is a posture choice, not a free upgrade. Some features that depend on cloud-hosted models behave differently:
We won't silently fall back to a cloud model. The UI tells you, the audit log records it, and your DPO can verify it.
Sovereign is an Enterprise feature and a 4-week onboarding — there is no self-serve path. The form below routes to our solutions engineering team; we usually reply within one business day.