“Legal and the DPO vetoed the cloud AI tool”
And the business still wants the capability — so the answer cannot stay “no”.
Mentis · Services · Sovereign AI architecture & MLOps
AI infrastructure your data never has to leave
Reference architectures for on-premise LLM inference and EU-hosted stacks, plus the pipelines, monitoring and retraining that keep models working — designed for organisations whose legal, compliance or clients rule out US clouds.
When to call us
Sound familiar?
“Our pilot runs on someone’s laptop”
It works, nobody can reproduce it, and it will break the day that person is on holiday.
“The cloud AI bill and the lock-in worry us”
Per-token costs climbing, data residency unclear, and no path back out.
What you get
Outcomes, not deliverables first
Sovereign by design
Inference, vectors, logs and models on your premises or in EU-only hosting — data residency as a fact, not a clause.
Right-sized
GPU sizing and open-weight model choice against your real load — serious capability without an overbuilt data centre.
Operable by your team
Pipelines, runbooks, monitoring and training so the stack is yours to run — or ours to run for you.
What we deliver
The scope you can actually sign
Target reference architecture
On-premise, EU-hosted or hybrid — components, data flows, security zones, the sovereignty rationale your DPO can sign.
Sizing & cost model
GPU/CPU sizing, model selection (open-weight LLMs, embeddings), licensing, three-year cost compared with cloud equivalents.
Inference stack deployment
LLM serving, vector database, API gateway, authentication — installed, secured and load-tested on your infrastructure.
MLOps pipelines
Training and evaluation pipelines, model registry, CI/CD for models and prompts, reproducible environments.
Monitoring, logging & audit trails
Performance, drift, cost and usage dashboards; the record-keeping the AI Act expects for high-risk systems.
Runbooks & enablement
Operations documentation, incident playbooks and hands-on training for your IT team.
How it runs
Fixed scope, fixed price per phase — you decide at each step.
- 1
Assess & design (2–3 weeks)
Constraints, workloads, existing estate — and a reference architecture with a cost model.
- 2
Build (4–8 weeks)
Stack installed, secured, integrated with identity and your systems, load-tested.
- 3
Harden & hand over
Monitoring, runbooks, training — your team operates, we stay on call.
Regulation & sovereignty
Why sovereignty is a compliance answer, not just a preference
GDPR transfer rules, DORA’s ICT third-party requirements in finance, professional secrecy in legal and healthcare, and the AI Act’s logging and record-keeping duties for high-risk systems all become simpler when inference, data and logs stay on infrastructure you control. We design the architecture so the compliance argument is in the diagram, not in a contract clause.
Questions
Is on-premise really viable for LLMs?
Yes, for most enterprise workloads. Open-weight models in the 7–70B range on one to a few GPUs serve assistants, extraction and summarisation for hundreds of users; we size against your real load and show the numbers.
How does the cost compare with cloud APIs?
It depends on volume and confidentiality value. We produce a three-year comparison — hardware, hosting, people — against cloud equivalents, so the decision is financial and regulatory, not ideological.
Can we go hybrid?
Often the best design: sensitive workloads on-premise or EU-hosted, non-sensitive ones on cloud APIs through EU regions, with routing rules that encode the policy.
We already run on Azure / AWS — does that disqualify us?
No. EU regions, private endpoints and customer-managed keys cover many cases; for the rest, a sovereign enclave for the sensitive workloads sits next to your existing estate.
Other services
Start here
AI Readiness Assessment
A 4-week, fixed-price diagnostic across six dimensions: maturity radar, scored opportunity portfolio, AI Act exposure register and a 90-day plan — the entry point for most clients.
Learn more →ServiceEU AI Act compliance & governance
Know exactly what the EU AI Act asks of you — and be able to prove it
Learn more →ServiceGenerative AI services
Generative AI that answers from your documents — inside your walls
Learn more →ServiceMachine learning & predictive modelling
Predictions you can explain, defend and act on
Learn more →ServiceOptimization & operations AI
Optimization that moves the P&L — routes, schedules, resources
Learn more →ServiceCustom AI application development
From a model that works to a product your people actually use
Learn more →Next step
Start with a 30-minute conversation
Tell us the problem, not a spec. You get an honest read on feasibility, data, compliance exposure and a first step — within one business day.