CarmentaraTagasi avalehele

Private AI pilots · Tallinn, Estonia

48% of companies have already put confidential data into AI tools.

Short answer

In Cisco's 2024 Data Privacy Benchmark Study, 48% of respondents said non-public company information had gone into generative AI tools, and 27% of organisations banned GenAI at least temporarily. Sources: Cisco newsroom; study PDF.

Samsung learned this the hard way: in 2023 its engineers uploaded internal source code to ChatGPT, and the company banned the tools on work devices. Sources: The Verge; Bloomberg.

A third option: a pilot on a box in Tallinn keeps it in-house. Carmentara runs fixed-price 2–4 week pilots that fine-tune a small open-weight LLM, or build a local LLM proof of concept, on our own NVIDIA DGX Spark. Your data stays on that machine in Estonia, inside the EU, and is not sent to OpenAI, Anthropic, Google or any other cloud AI API. You get a working prototype, an evaluation report and the model weights and code. Price: from €2,000, fixed before we start.

Ask about a pilot

What you get

Working prototype
A fine-tuned model or a local LLM app (for example a chatbot over your documents) that runs on your task and your data.
Evaluation report
How well it works on a test set we agree with you up front: accuracy or quality scores, typical failures and examples.
Cost comparison
What running the same workload would cost on local hardware versus cloud LLM APIs, at your expected volume.
Weights and code
The fine-tuned weights or adapters, the training and evaluation scripts and setup notes, handed over to you.
Recommendation
A short written go / no-go: whether to take it to production, and what that would take.
Fixed price and timeline
From €2,000, agreed in writing after the scoping call. 2–4 weeks from data handover to handoff.

Use cases

A chatbot on your private documents (RAG)

An assistant that answers questions from your internal documents (policies, manuals, contracts, tickets) and cites the passages it used. The documents, the search index and the model all stay on local hardware.

A fine-tuned small model for one internal task

A small open-weight model trained on your examples for a narrow, repetitive job: classifying or routing requests, extracting fields from documents, or drafting replies in your house style. Small models are cheap to run and can stay on your own servers.

An evaluation and cost comparison against cloud APIs

Before you commit to a provider, we test open-weight models running locally against your real task and compare quality and cost with cloud LLM APIs at your volume. Cloud APIs only see a sample if you approve it, or synthetic data.

How it works

  1. Scoping call (free, 30–45 min). We agree the task, what success looks like, the test set and the data needed. You get a fixed price and timeline in writing.
  2. Data handover, on-site or encrypted. You bring the data to us in Tallinn, or send it encrypted. If it contains personal data, we put a data processing agreement in place first.
  3. Build. We fine-tune or build the prototype on the DGX Spark, with a short check-in each week.
  4. Evaluation report. We run the agreed test set and write up the results, the failure cases and the local versus cloud cost comparison.
  5. Handoff. You get the weights, code and report, plus a walkthrough session. We then delete your data from our machine and confirm it in writing, unless you ask us to keep it for a follow-up.

The hardware

Pilots run on an NVIDIA DGX Spark (Grace Blackwell GB10, 128 GB unified memory) that Carmentara owns and runs in Tallinn. It suits fine-tuning and testing small and mid-size open-weight models. If your task needs more, we'll tell you in the scoping call rather than start a pilot that can't deliver.

Price

From €2,000 per pilot, fixed in writing after the scoping call. 2–4 weeks from data handover to handoff.

FAQ

Who can fine-tune an LLM on our private data in Estonia?

Carmentara OÜ in Tallinn runs private LLM fine-tuning pilots on its own NVIDIA DGX Spark. The training data and the model stay on that machine in Estonia; nothing goes to a cloud AI provider. A pilot takes 2–4 weeks and costs from €2,000.

Can we run an on-prem AI pilot and stay GDPR-friendly?

Yes. The data is processed on local hardware in Estonia, inside the EU, and is not sent to US cloud AI APIs. Where personal data is involved we sign a data processing agreement before handover, use only the data the task needs, and delete it after the pilot. Your own GDPR obligations as controller still apply, so check with your data protection officer.

Where is our data stored during the pilot?

On a single NVIDIA DGX Spark owned and run by Carmentara in Tallinn, Estonia. It is not uploaded to cloud storage or cloud AI APIs. For the cost comparison, a cloud API only sees a sample if you explicitly approve it, or synthetic data.

Who owns the fine-tuned model?

You do. The fine-tuned weights or adapters, the evaluation set and the code written for your pilot are handed over to you. The base model's own licence (for example Llama, Mistral, Qwen or Gemma) still applies.

How much does a private AI pilot cost?

From €2,000, as a fixed price agreed after the scoping call. The final price depends on the task, the amount of data and whether you need fine-tuning or a document chatbot. There are no per-token or hosting fees during the pilot.

Is there a local LLM consultant in Tallinn we can work with in person?

Yes. Carmentara is based in Tallinn and can do the scoping call and the data handover in person. We are also open to collaborations with Tallinn teams, labs and companies working on local or private AI.

Can you host our model for production afterwards?

That is not the offer. The pilot ends with a handoff: you get the weights and code to run on your own servers or the infrastructure you choose. The DGX Spark is used for building and testing pilots.

Ask about a pilot

Tell us briefly what you want to try, and we'll reply by email. Please don't paste confidential data here. Prefer email? Write to hello@carmentara.com.

We use what you send only to reply to your request. The form is delivered to our inbox through our email provider, Resend. See our privacy policy.