Your own AI model, trained on your work.
We don't resell AI. We build it.
Insurance claims, contracts, discharge letters, customer questions: the work your team repeats every day, taught to a model that runs on your servers. The model is yours, your data never leaves the building, and it gets better with every cycle. We can start with a single decision.
30 minutes with the founder, who leads every project. No sales reps.
One repetitive task, data that can't leave the building
Pulls data from claims, contracts and reports into your fields. Hands doubtful cases to the claims handler.
Codes discharge letters to ICD-10, summarises and translates medical text. Data stays inside the hospital.
Anonymises, classifies and routes documents by your rules, with a log of every step.
Answers from your knowledge base, in your language, and cites the source. Sends only after a person approves.
From one decision to an agent
Most companies start with one decision that repeats a hundred times a day. Once that works, we go deeper.
One decision, with a probability
Where this e-mail goes, what this alarm is, whether to pay this invoice. A small model picks only among the allowed options and says how sure it is. Below the threshold a person decides.
Writes like your team
A model trained on your documents and answers: your terms, your format, your tasks. You get the weights; it runs on your servers.
Runs a whole process
Reads the inputs, calls your systems, drafts, and waits for a named person where the rules say so. With a log of every step.
First we measure whether it pays off
We don't sell projects on trust. One day, fifty of your cases and a written verdict before you commit to anything bigger. Every price is on the page.
Evaluate
Fifty of your real cases through several models. A written verdict on whether a model or an agent pays off, and what it would cost. NDA first.
Pilot
One task, an agreed target number, a fixed price. A model on your data from €3,200, an agent for one process from €6,900.
Deploy
On your servers, an air-gapped machine or EU hosting. You decide who approves what; every step is logged.
Keep it current
New data every month, retraining and a fresh measurement. When a better open base appears, we move to it within days.
Pricing and details → · Own model or API? Calculator → · Security and compliance: NDA, DPA, data in the EU →
A small company that only needs EU AI Act compliance? AI use review from €490 fixed → · Working in Croatian, Serbian, a Baltic or another under-served language? Your language, your model →
A model that learns from its own work
Most models are at their best on the day they are delivered. Ours get better: every decision, every correction and every uncertain case becomes training data for the next version.
- WorkA large model or a person starts. They keep handling the uncertain cases the small model cannot do yet.
- LogEvery decision and every correction is logged. A person's approval is also a training example with the right answer.
- SimulatorWhere real cases are too few, a simulator creates them, with a known correct outcome.
- TrainingFrom this we train a smaller model that does only this task: faster, cheaper and with less room for abuse.
- MeasurementA new version goes live only if it beats the previous one on the frozen set. A person approves the swap.
On one of our models such cycles raised the share of correct decisions not to call a tool from 0.847 to 0.956 on 2,471 held-out steps.
That is the retainer: Watch measures every month, a cycle trains when needed. Pricing →
We run our own agents first
Everything we offer is something we already depend on. Eight systems run on the same stack we would build for you.
An agent that checks every price and deadline, every day
Every day it reviews all built pages: titles, links, structured data, duplicate content. It keeps a facts register and checks each price against its source; anything without a source does not stay on the page. Every fix is prepared as a proposal and published by a person.
For you: price lists, terms and tender documents that are always correct.
Common questions
When we are not the right choice
If a general chatbot on public information is enough, a hosted model from a large provider costs less than training one. And if you need legal advice on the AI Act, talk to a lawyer: our review is technical documentation, not legal advice.
What does MediaAtlas do?
MediaAtlas is an AI research and development studio in Sevnica, Slovenia. Most of our work today: we train language models on a company's own data and install them where that data lives, on the company's servers or with EU hosting. The customer gets the weights. We also build agents on those models and run fixed-price EU AI Act reviews.
Is MediaAtlas the same company as Atlas Media?
No. MediaAtlas d.o.o. is a Slovenian company based in Sevnica, founded in 2012, written as one word. It is not connected to businesses called Atlas Media.
Is MediaAtlas a motion capture studio?
Not only. MediaAtlas is an AI research and development studio. Motion capture and real-time 3D are where we started, and the studio has its own page. Since then we have built a range of other products: cold-chain monitoring (Sensoram), working-time records (EDC), verifiable data (SemantiCord), hydrogen trading (EnergonX) and coding for children (Mali Pokovci). Today our main work is training language models on company data and the AI agents that run on them.
What does it cost to start?
A one-day evaluation on fifty of your real cases costs €1,900 excluding VAT and is credited against a pilot. Pilots start at €3,200. An AI Act review for a company of 10 to 50 people costs €990.
Where can I check your results?
Our models, model cards and evaluation numbers are public on Hugging Face under MediaAtlas and texdata. Code for our discovery agent is on GitHub.
What we learn, written down
Practical notes from our own work: regulation, training, evaluation. All posts →
A small fine-tuned model for customer support in Slovenian: accuracy, frontier fallback, cost
Which tickets suit a model, how many past replies you need, a 4B–12B model with retrieval and a frontier fallback, how we measure it and what it costs.
BlogAnonymising Slovenian documents with a local model: accuracy, GDPR and cost
Why regex fails on Slovenian names and EMŠO, how a local 4B–12B model plus rules is measured for recall, what GDPR counts as anonymous, and what it costs.
Products we built
Two SaaS products for regulated work in Slovenia, in use by customers, and projects from verifiable data to hydrogen and robotics for children. They are why we know what an audit asks for and how a product survives real use.
Sensoram
Real-time temperature and humidity for pharmacies, hospitals, labs, food and logistics. An alarm before it is too late, and a report ready for the inspector.
- EN 12830
- hash-chained records
- escalating alarms
- SMS, Telegram, e-mail
EDC
ZEPDSV-A records for Slovenian employers: web, offline mobile app, NFC and QR clock-in. Breaks, night work and overtime caps are calculated automatically.
- ZEPDSV-A
- NFC and QR
- payroll export
- period locking
SemantiCord
Memory Units: cryptographically signed JSON envelopes that prove where data came from and that it has not changed, in the age of deepfakes.
- RFC 8785
- multi-party signatures
- schema registry
EnergonX
A location-aware hydrogen trading platform: certified origin, integrated logistics and transparent tracking from producer to consumer.
- green, blue, grey H₂
- certified origin
- logistics
AssetManiac
Several AI agents read on-chain data, market flows and news at once, cross-check each other and condense it into one clear insight.
- multi-agent
- Telegram and API
- 0–100 scoring
Mali Pokovci
From bricks to code: children aged 6 to 14 build LEGO robots and program them. A mistake is a signal, not a failure, and they work in pairs.
- ages 6–14
- LEGO robotics
- visual programming
Motion Capture Studio
Motion capture for video games, VR and film: real-time animation that brings characters to life. It is where we started.
- games and VR
- film
- real time
Bring us one process.
The job your team repeats most. We'll tell you, with numbers, whether AI should do it, and what it would take.