Your data stays. The AI comes to it.

Calexis Labs sets up AI for companies: on your own servers, in the cloud, or both. When an off-the-shelf model is not good enough, we train one on your work.

We took an 8B open model from 17 to 102 correct answers on 150 network faults it had never seen. It runs on one machine.

  • Runs on your hardware
  • EU company, Tallinn
  • Open source CLI and weights

What we set up for you.

Every engagement starts with a two-week assessment. You get a written answer before you commit to anything bigger.

Local LLM infrastructure

Open models running on your own servers, sized for your load and connected to your documents.

  • GPU sizing and serving
  • Access control and usage logs
  • Search over your own documents

Cloud LLM architecture

Claude, GPT or Gemini, set up safely for every team in the company.

  • One gateway for all teams
  • EU data region
  • Cost limits per team

Custom model training

When a general model is not good enough, we train a small one on your task and prove it with an exam.

  • Trained on your own cases
  • Exam on questions it never saw
  • Runs fully on your hardware

Run and improve

We keep it running after launch, on your hardware or in your cloud account.

  • Monitoring and updates
  • Monthly cost report
  • New scores with every release
Case study

Kestrel: a small model, trained for one job.

Network engineers lose hours finding the cause of an outage. We trained Kestrel, a 5 GB model, to diagnose network faults inside closed networks, then graded it on faults it had never seen.

100%75%50%25%0%
11%
68%
85%
Qwen3 8B (base)Fully local · ~5 GB17 of 150 correct
Kestrel v1our modelFully local · ~5 GB102 of 150 correct
Claude Sonnet (reference)Cloud reference model128 of 150 correct

These are controlled lab results, not measured production accuracy. The Claude Sonnet reference includes six infrastructure errors in the 150-question denominator. Context settings differ between base Qwen and Kestrel.

How we work.

  1. 0130 minutes

    Call

    We ask about the task, the data and where it is allowed to go.

  2. 022 weeks

    Assessment

    A written answer: local, cloud or both, what it costs and how well it will work.

  3. 032 to 10 weeks

    Build

    We set up the infrastructure or train the model, then test it on cases it has never seen.

  4. 04Monthly

    Run

    We keep it running and improving. Every release comes with new scores.

Why work with us.

Private by default

Sensitive work can stay inside your network. Nothing has to leave it.

Predictable cost

Your own hardware, or cloud usage with a hard limit per team. No surprise bills.

Measured, not promised

Every model we ship comes with an exam score, not a demo.

We run our own products too.

We don't only build for clients. Our own software is live and paid for, so we know what it takes to keep it working.

Questions we hear first.

What hardware do we need for a local model?

For a small model like Kestrel, often a single server with one modern GPU. The assessment tells you exactly what to buy or reuse, before you spend anything.

Does our data leave our network?

Only if you decide it can. Sensitive work can run fully on your servers. Cloud models are used only for the tasks you allow, in an EU region.

How do we know it actually works?

We build an exam from your own cases and report the score, the way we did for Kestrel. You see the number before you decide to go further.

What do we get from the assessment?

A written report: which tasks fit, local, cloud or both, the hardware and running cost, and an honest go or no-go. It is yours to keep, even if you stop there.

Is there a task your team repeats every day?

Tell us about it. We'll tell you honestly if a model can take it over. We usually reply within a day.

hello@calexislabs.com