Built for the questions your CISO will ask.
Large organisations don't buy a demo. They buy guarantees they can write into a contract. Here are ours.
Sovereign
Servers in the European Union, under European law. No hyperscaler and no US AI API in the default path, so no exposure to extraterritorial laws such as the US CLOUD Act.
Sealed
Each client gets its own memory, its own access keys and its own logs. One client's data can't reach another's, by construction rather than by policy.
Fast
Dedicated GPUs that answer only to our clients. No queue behind the rest of the world, no rate limit, no outage you can't do anything about.
Portable
Your knowledge lives in a harness that belongs to you. Change the model tomorrow: nothing to migrate, nothing lost.
One door in. No way around it.
Three zones, each with a single job. Your users only ever reach the first one. The models sit at the back, with no public endpoint of their own.
The only way in
European data centre · EU- TLS encryption on every connection
- Authentication and per-project access keys
- Strict security headers, no third-party scripts
Your product and your memory
European data centre · EU- Your application and its database
- The harness: your vector memory, one collection per client
- Nightly backups, with an automatic restore test
The models, out of reach
European data centre · EU- Dedicated GPU servers we operate
- The latest open-source models, run locally
- No public endpoint: reachable only through the encrypted link
Dedicated servers in top-tier European data centres, operated by our team. No hyperscaler and no US AI API in the processing chain.
Every piece of data, on the map.
No vague “secure cloud”. Here is what we store, where, who can see it, and for how long.
| Data | Where | Who can see it | How long |
|---|---|---|---|
| Conversations with the AI coach | Application database, EU | The user. Never the employer, except aggregated and anonymised trends, with users' knowledge | As long as the account exists, or the period set in your contract |
| Your content and methods | Harness (vector memory), EU | Your users, through the product. Your client space only | For the duration of the contract, then returned and deleted |
| What the model receives | Our GPU servers, EU | No one: processed in memory to produce the answer | Not kept once the answer is produced |
| Voice recordings, scanned documents | Our GPU servers, EU, for transcription or OCR | No one: only the resulting text is kept | Source file deleted after processing, unless your product needs it |
| Technical logs | Our servers, EU | Our engineering team, for security and support | Limited retention, set in your contract |
Sealed tight, client by client.
One memory per client
Each client's knowledge sits in its own collections, named after the client. A request that tries to reach another client's space is refused before it touches the data.
Keys that open one door
Every project has its own access key, valid only for its own data. Only our platform key can perform administrative operations.
Encrypted in transit, everywhere
TLS between the browser and the gateway, an encrypted link between the application and the models. Nothing travels in clear text.
No training, no resale
Your data is never used to train a model, ours or anyone else's, and never sold. The models we run are fixed; they don't learn from your users.
Erasure in one call
When a user or a client leaves, their entire memory is deleted in a single operation. Right to erasure (GDPR, article 17), applied for real.
Least privilege, by default
Services run without administrator rights, in separate containers. The site you're reading loads no third-party script, font or tracker.
Local means fast.
Our GPUs serve only our clients. No shared queue with millions of users, no rate limit on busy days, no surprise when a provider changes its terms. The answer starts before your user has finished reading the question.
Measured on our infrastructure in October 2026, with a latest-generation open-source model, averaged over several runs. Figures vary with the model and the load; we measure them again for your project.
When the answer is on a list, the AI picks. Instantly.
Most of what an organisation asks an AI isn't prose. It's a decision: which category, which team, which priority, yes or no. For those, our models don't write a paragraph. They choose from your list, and only from your list.
Always a valid answer
The output is constrained to your options at the source. No free text to parse, no invented category to catch.
Confidence you can measure
Each choice comes with the model's real probability, not a score it writes for itself. Below your threshold, a human takes over.
Every decision on record
Same question, same answer: decisions are logged and reused, with the model and the probability behind them. An audit trail by design.
Sort incoming documents · Qualify a lead · Route a request to the right team · Flag a risk · Tag a conversation
Your knowledge, whatever the model.
A model is a tool. The harness is your asset: your content, methods, history and constraints, in a memory that belongs to you. Every model plugs into it; none of them owns it.
Your method, owned by you.
Technically, a vector (RAG) memory of your data and past interactions, independent of the models that use it. Portable from one model to the next.
The open-source building blocks we run locally
Source: anonymised client case, over one year on our infrastructure.
And if you want a US cloud model for one specific task, that's your call. The harness stays with you, and you can switch back whenever you like.
Three ways to run it. You choose.
On our European servers
The fastest way to start. Your dedicated client space on our infrastructure in the European Union, operated by us.
- Live in weeks, not months
- Hosting, updates and monitoring included
- Fixed monthly cost
On your own servers
For the most sensitive data. We install the whole stack, models included, in your data centre or your private cloud.
- Your data never leaves your premises
- Can run without internet access
- Your teams keep the keys
Hybrid
The harness and sensitive data stay with you; heavy computing runs on our European GPUs, through an encrypted link.
- Your memory stays at home
- GPU power without the hardware
- Switch to option 2 at any time
The questions we answer in writing.
These are the answers we put in your contract. Ask for the details: we'll go through them with your teams.
Is our data transferred outside the European Union?
No. Application, database, memory and models all run in EU data centres. No US provider sits in the processing chain.
Who are your sub-processors?
Only European providers: the data centres hosting our dedicated servers, and a European service for transactional emails. The full list is written into your contract, and any change is notified to you in advance.
Do you sign a data processing agreement?
Yes. A data processing agreement under article 28 of the GDPR is part of every contract, with the security measures described on this page.
Is our data used to train models?
Never. The models we run don't learn from your users, and your data isn't shared with any model provider.
Can we get our data back?
At any time. Your documents, your conversations and your harness are exported in open formats. The code we write for you is yours too.
What happens if Unfollow Studio disappears?
You keep everything: code, harness, data and models are open and documented. Another team, or yours, can take over without starting from scratch.
Can we audit the infrastructure?
Yes. We walk your security teams through the architecture, the access controls and the logs, and answer your security questionnaires.
Sovereign AI,
on your terms.
Let's talk.
Tell us about your data, your constraints and your users. Our CTO will answer you with a concrete architecture, not a brochure.
Based in Lisbon · Working across Europe