Platform architecture
The Sovereign Control Plane for Enterprise AI.
One gateway between your applications and every model provider. It decides where each prompt may run, strips what must not leave, routes to a certified in-region endpoint and returns a signed receipt you can hand to a regulator.
Last decision / eu-central
tenant tnt_de_4417 domicile DE
policy eu-gdpr-strict@4.2.0 ALLOW
pii PERSON, IBAN masked
route fr-par / SCW-DC5
us-east-1 excluded: zone US
overhead 6.1 ms receipt leaf 18842
A single chat completion from a German insurer's claims assistant, followed through all seven stages. Each stage runs in the EU data plane in memory; the only things persisted are the decision log entry and the signed receipt.
Stage 1 of 7 / Your service, unchanged SDK
Client App
Your backend calls the OpenAI-compatible chat completions API. The only changes are the base URL and two headers that name the policy pack and the tenant.
POST https://eu.yttsecunet.com/v1/chat/completions
Authorization: Bearer jr_live_************************
X-YTT-Policy: eu-gdpr-strict@4.2.0
X-YTT-Tenant: tnt_de_4417
{
"model": "mistral-large",
"messages": [{
"role": "user",
"content": "Kunde Jonas Becker, IBAN DE89 3704 0044 0532 0130 00, fragt nach dem Status seiner Erstattung."
}]
}Four pillars, one request path
Jurisdiction-Aware Inference Router
Resolves every request to an endpoint inside the legal zone its policy allows, using tenant domicile, user geo-IP and locale, or data class. Failover stays inside the zone: Frankfurt falls back to Amsterdam or Dublin, never to Virginia.
tnt_de_4417 DE -> zone EU geo 194.25.x DE -> zone EU class health -> HDS pool resolved -> fr-parOpen the deep dive
Policy-as-Code Compliance Engine
Open Policy Agent evaluates Rego packs for GDPR Chapter V, UK GDPR, Swiss revFADP, the EU AI Act and Gulf residency laws. Every decision is logged with its input hash and pack version.
deny contains msg if {
not input.endpoint.zone in allowed
}Open the deep dive Zero-Egress PII & Cross-Border Guardrail
Names, IBANs, passport numbers and national identifiers such as the German Steuer-ID or UK NHS number are masked or pseudonymized in memory. The reversible mapping stays in a vault in the origin region.
Jonas Becker -> <PERSON_7f2a> DE89 3704 0044 ... -> <IBAN_c91e>Open the deep dive
Cryptographic Proof-of-Residency Ledger
Each response carries an Ed25519-signed JWS receipt naming the endpoint, provider region, facility and upstream certificate thumbprint. Receipts are SHA-256 leaves in an hourly Merkle tree that anyone can verify offline.
alg EdDSA region fr-par facility SCW-DC5 leaf 18842 / root 3e0b...a41c VERIFIEDOpen the deep dive
Buyers in privacy and security teams ask us the same question first: where does your responsibility stop? Here is the plain list, because a residency control that overstates itself is worse than none.
- We do not host foundation models. YTT Secunet is the sovereign routing and governance layer. It orchestrates certified provider endpoints and your self-hosted runtimes, and evidences where each request ran.
- We do not give legal advice. Policy packs encode common readings of GDPR Chapter V, the AI Act and national rules. Your counsel decides which pack applies and what exceptions you accept.
- We do not replace your DPIA or transfer impact assessment. We supply the technical evidence those documents need (routes, decisions, receipts), but the assessment and its conclusions remain the controller's.
- We do not make a US-parented provider immune from the CLOUD Act. We reduce exposure with zero prompt retention, customer-held keys and in-VPC deployment, and we label each endpoint's parent jurisdiction so you can choose.
- We do not store prompts or completions. They stream through memory and are never written to disk or used for training. Only metadata, decisions and receipts are kept.
What we do own: the decision for every request, the enforcement of that decision in the data path, and a signed record of the outcome that does not depend on trusting us. If a regulator asks where a given inference ran, the answer is a receipt and an inclusion proof, not a vendor letter.
- < 8 ms
- p99 added gateway latency, policy evaluation plus receipt minting
- YTT Secunet production telemetry, 2026
- < 3 ms
- median OPA policy evaluation per request
- YTT Secunet production telemetry, 2026
- 80 to 100 ms
- typical round trip between Frankfurt and US East (Northern Virginia) before any model compute
- Speed-of-light fibre path, public cloud inter-region latency measurements
- EUR 20M or 4%
- of worldwide annual turnover, whichever is higher: GDPR Article 83(5) ceiling for unlawful international transfers
- Regulation (EU) 2016/679, Art. 83(5)(c)
YTT Secunet routes to endpoints that pass registry certification. Each card names the deployment we target, not the provider as a whole: a provider's US regions are separate endpoints with a separate zone.
Azure EU Data Boundary
Azure OpenAI deployments in EU regions such as Sweden Central and Germany West Central, used with EU Data Zone deployment types.
AWS European Sovereign Cloud
AWS's separate EU cloud, operated by EU-resident staff, with its first region in Brandenburg, Germany.
Google Cloud Sovereign Controls
Vertex AI in EU regions with Assured Workloads data boundary and residency controls.
Scaleway
French provider with Paris, Amsterdam and Warsaw regions and hosted open-weight models.
OVHcloud
French-headquartered cloud with AI Endpoints for open-weight models in European data centres.
Swisscom
Swiss-operated infrastructure for workloads that must stay in Switzerland under revFADP and FINMA outsourcing rules.
STACKIT
The Schwarz Group's German cloud, with data centres in Germany and Austria and hosted model serving.
On-prem Kubernetes vLLM
Your own GPU nodes running vLLM's OpenAI-compatible server, registered as a private sovereign endpoint.
The gateway speaks the OpenAI and Anthropic wire formats, so existing SDKs, LangChain chains and agent frameworks keep working. You change the base URL to the regional entry point for the tenant (eu, uk, ch, gcc, apac or us) and add two headers.
The model name you send is a logical alias. The registry maps it to a concrete deployment inside the allowed zone, so the same code serves a German tenant from Paris and a Swiss tenant from Zurich without branching.
- Streaming, tool calls and JSON mode pass through unchanged
- Receipt header on every response, including streamed ones (sent as a trailer)
- Policy denials return HTTP 451 with the decision id, never a silent reroute
import os
from openai import OpenAI
client = OpenAI(
base_url="https://eu.yttsecunet.com/v1",
api_key=os.environ["YTT_SECUNET_API_KEY"],
default_headers={
"X-YTT-Policy": "eu-gdpr-strict@4.2.0",
"X-YTT-Tenant": tenant.id,
},
)
resp = client.chat.completions.with_raw_response.create(
model="mistral-large",
messages=[{"role": "user", "content": prompt}],
)
receipt = resp.headers["x-ytt-proof-signature"]
completion = resp.parse()Does YTT Secunet host the language models?
No. YTT Secunet is the routing and governance layer. Inference runs on certified provider endpoints (for example a provider's EU regional deployment) or on your own vLLM or Ollama runtime. We certify, route to and evidence those endpoints; we do not train or serve foundation models.
How much latency does the gateway add?
YTT Secunet measures under 8 ms p99 added latency for policy evaluation, PII detection and receipt minting in its own production telemetry. For a European user the bigger effect is usually the route itself: a Frankfurt to US East round trip is roughly 80 to 100 ms before any model compute, which an in-region endpoint avoids.
What changes in our application code?
The base URL of your OpenAI or Anthropic compatible SDK, plus two headers: X-YTT-Policy names the policy pack and version, X-YTT-Tenant names the tenant whose domicile drives routing. Request and response bodies keep the provider's schema.
Where is the control plane hosted?
The multi-tenant control plane runs in Frankfurt am Main with a standby in Paris, on EU-headquartered infrastructure. Regional data planes run in the EU, UK, Switzerland, the GCC, APAC and the US. Sovereign Dedicated places both planes in your VPC or on-prem, including air-gapped sites.
Does using YTT Secunet make our AI processing GDPR compliant?
It enforces the technical part: where inference runs, what personal data leaves the region and what evidence exists. You still need a lawful basis, a DPIA where Article 35 requires one, and a transfer impact assessment for any transfer you choose to allow. YTT Secunet does not give legal advice.
Does it remove CLOUD Act exposure?
It reduces it by design. Prompts are not retained, so there is no stored content to disclose; signing keys can be customer-held; and Dedicated runs in your own environment. It cannot change the legal position of a US-parented model provider, which is why the endpoint registry records each provider's parent jurisdiction.
Sovereign route ready
Stop risking 4% global turnover GDPR fines on AI inference.
Deploy sovereign, cryptographically proven AI routing in 15 minutes.