For EU companies, public-sector buyers, and regulated industries

Build on the AI stack
you can defend in front of a regulator.

Sovereign-first AI deployments on French and EU-resident clouds — open-weight model routing, full audit trail, EU AI Act-ready from day one. For buyers where data residency, procurement constraints, or AI Act exposure are non-negotiable.

Live engagement: sovereign AI infra for a French construction group
EU AI Act risk-classified · ANSSI-friendly architecture
Procurement-friendly pricing, FX-neutral

US-only AI is a procurement risk and a compliance risk.

Most European companies want the AI velocity but not the vendor lock-in, the FX exposure, or the regulator visit.

01

Procurement disqualification

Your customer's procurement team asks for EU residency. Your AI platform sits in us-east-1 with a US-headquartered provider. The deal stalls in legal review for 8 weeks.

02

EU AI Act high-risk classification

You did not realise your pricing model, your screening tool, or your HR co-pilot is high-risk under Annex III — until the conformity assessment is due in 90 days.

03

Single-vendor lock-in

A pricing change at one US provider breaks your unit economics overnight. You have no model-routing layer, no fall-back, and your stack speaks one API.

A sovereign-first stack with frontier-quality output.

We route across EU clouds, open-weight models, and frontier providers (with EU residency where it matters) so you get the velocity without the lock-in.

  • Sovereign EU clouds, by default Scaleway, OVHcloud, Outscale for compute and data residency. Public-sector and regulated buyers stop asking awkward questions in procurement.
  • Open-weight models, when sensitive Mistral, Llama, Qwen self-hosted on EU infrastructure for the workloads where data cannot leave. Frontier providers for the rest.
  • EU AI Act conformity, baked in Risk classification, technical documentation, post-market monitoring, audit trail — designed in from day one rather than retro-fitted before an audit.
  • Hyperscaler when warranted AWS, Azure, GCP with EU residency, Bedrock or Azure OpenAI for clients already on a hyperscaler — we are not religious about sovereignty for the sake of it.
Stack we route across
EU sovereign cloudsScaleway · OVHcloud · Outscale
Open-weight modelsMistral · Llama · Qwen · Hugging Face
Frontier (EU residency)Anthropic Claude · OpenAI · Google
Hyperscaler (EU region)AWS · GCP · Azure with eu-* regions
Data layerPostgres pgvector · Qdrant · Supabase EU
EU AI Act toolingRisk register · Annex IV docs · monitoring

The sovereign routing console, live.

A representative snapshot of our sovereignty engine. Click tabs, switch language — everything is built to enforce data residency at the edge, not in policy decks.

Insightrix · Sovereign Routing Console Live Routing Demo
Aaru.eu

Routing engine

Classifier → sovereignty rules → model. Cited every step.
Live traffic last 24h
01 InputDocument · query1,284 req/min
02 ClassifierSensitivity · PII12ms median
03 PolicyEU-only · no US routeenforced
04 Modelmistral-largescaleway-paris
Requests
1,284
+18%
p50 latency
142ms
-9ms
Cost / 1k req
€1.42
+3%
EU-routed
100%
0 US egress
Traffic by hour (last 24h)
Peaks coincide with European business hours · minimal night-time load
Model mix
62%
mistral-large · 62%
claude-eu · 24%
llama-fr-self · 14%

Data residency map

Every workload pinned to a sovereignty class. Enforced by the policy engine, not by trust.
0 cross-border events / 30d
All workloads compliantNo US-region routing detected in the last 30 days. Failover chain remains entirely EU-resident.
WorkloadData classRegionModelStatus
sensitive_workloadsPII · client docsscaleway-parismistral-largeEU
general_reasoningInternalovhcloud-gravelinesclaude-sonnetEU
vision_docsOCR · scansovhcloud-gravelinesclaude-visionEU
offline_evalEval suiteon-prem · parisllama-fr-selfEU
public_searchPublic webscaleway-parismistral-smallEU

Audit log · routing events

Every routing decision captured with a cited policy. Exported nightly to your SIEM.
last 30 min all severities
14:42:18OKRoute resolved · req_8af2mistral-large@scaleway-paris
14:42:11INFOClassifier confidence 0.96 · flagged as sensitive_workloads
14:41:54OKFailover ready · claude-eu · llama-fr-self
14:41:32WARNPolicy reject · US endpoint blocked · rerouted EU
14:41:12INFODrift check · hallucination rate 0.4% (under threshold)
14:40:48OKSnapshot exported · siem://insightrix/2026-05-17
Recent policy events
14:42Policy v3.2 applied to all sensitive workloads
13:08Model mistral-large promoted to primary for FR traffic
09:20Nightly export pushed to SIEM · 14,221 events
ydayFailover chain rotated · llama-fr-self added as L3
Export targets
siem-splunk · nightlyactive
s3://eu-west-3/auditactive
elastic-eu · realtimeactive
DPO mailbox · weekly digestMon 09:00

Model swap · config diff

From US-region Anthropic to EU-region Mistral. One PR. Audited and reversible.
routing.yaml +3 −1
infra/routing.yaml feature/eu-failover · reviewed by @aru
1sensitive_workloads: claude-3.5-sonnet # us-east-1
1+sensitive_workloads: mistral-large # scaleway-paris
2 general_reasoning: claude-sonnet # eu-residency
3 vision_docs: claude-vision # ovhcloud-gravelines
4+failover_chain: [mistral, claude-eu, llama-fr-self]
5+policy_engine: "eu-only" # enforced at edge
Why this swap?US-region endpoint flagged by procurement · EU AI Act Art. 27 requires data residency proof. Mistral large benched within 4% on the IFRS-RAG suite — acceptable trade-off.
Quality delta (eval suite)
Recall @50.94 → 0.92
Citation accuracy0.97 → 0.96
p50 latency142ms
Routing engine 1.4.0 · 14 models0 US egress todayp50 142msp99 412msscaleway-paris · ovhcloud-gravelines live req 142,481

What our engagements typically deliver.

100% EU-resident infrastructure where required
4–8 weeks kick-off to production
EU AI Act conformity documentation, day one

Tell us where your sovereignty constraint sits.

A 30-minute call. We will look at your buyers, your data, your regulator, and your existing stack — and tell you what 'sovereign AI' actually has to mean for you.

  • We will tell you when sovereignty is overkill for your use case.
  • NDAs available before the call.
  • You'll leave with a one-page architecture proposal.
Email aru.bhardwaj@insightrix.eu directly — founder, based in Paris.

By submitting, you agree to our privacy policy. We respond within 24 working hours and never share your details.

Common questions before we talk.

Does sovereign AI mean Mistral-only?

No, and we would not recommend it. Mistral is great for many workloads; for others Anthropic Claude or open-weight Llama is the right call. Sovereignty is about data residency, audit trails, and procurement — not about religious model choice.

How do you handle clients with US, UK, and EU subsidiaries?

Routing layer. The same application talks to a US-resident endpoint for the US subsidiary and an EU-resident endpoint for the EU subsidiary, with policy enforced at the routing layer. Your subsidiaries each see the cloud their regulator expects.

Are you a Scaleway / OVHcloud / Mistral partner?

We are not in their formal partner programmes. We deploy on these providers because they are the right answer for sovereign-first buyers; we have no commission incentive bending our recommendations.

What does it cost?

Every engagement is scoped after a 30-minute discovery call and a walk-through of your data-residency and procurement constraints. We send a written, fixed-scope proposal — with timeline, EU AI Act documentation included, and a transparent quote — within 5 working days. Public-sector procurement-ready packaging on request.

Ship sovereign-first AI without sacrificing velocity.

Book a 30-minute call. We will tell you what sovereignty actually has to mean for your buyer, your regulator, and your stack — and what you can ignore.

Book a Free Call