MediaAtlas is an AI research and development studio in Sevnica, Slovenia. We train language models on a company's own data and build agents on them, for businesses in regulated industries and under-served languages that need AI they own and can verify.
We don't resell AI. We build it.
One team takes a model from raw company documents to a measured, deployed system. Every service below is also sold on its own.
We fine-tune open-weight models (4B to 35B MoE) on your manuals, tickets, contracts and terminology, for chat, extraction, classification, translation and tool calling. The outcome is a model that does your task at the level of a large API, runs on your hardware, and belongs to you. Details
We turn raw material into training data: cleaning, deduplication, quality scoring, licence tracking, teacher distillation and domain translation with frozen terminology. What you don't have, we build, so a project does not stall on missing data.
We build a frozen benchmark from your documents, measure untuned models first, and score every version against it. Sold on its own as a one-day evaluation that tells you, in writing, whether training is worth it at all.
Retrieval over your documents with hybrid search, reranking and cited answers, and graph-based retrieval where relations matter more than passages. The outcome is answers you can trace back to a source, with access control per document.
Ingestion, OCR, anonymisation and pseudonymisation, extraction into schemas and validation, plus agents with native tool calling. Small-model-first cascades with a frontier fallback keep cost down without giving up quality.
Slovenian speech recognition and synthesis, fine-tuned the same way as text models. Everything we ship runs from a laptop to a multi-GPU server or an air-gapped machine, or EU-hosted if you prefer.
Products
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.
ZEPDSV-A records for Slovenian employers: web, offline mobile app, NFC and QR clock-in. Breaks, night work and overtime caps are calculated automatically.
Memory Units: cryptographically signed JSON envelopes that prove where data came from and that it has not changed, in the age of deepfakes.
A location-aware hydrogen trading platform: certified origin, integrated logistics and transparent tracking from producer to consumer.
Several AI agents read on-chain data, market flows and news at once, cross-check each other and condense it into one clear insight.
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.
Motion capture for video games, VR and film: real-time animation that brings characters to life. It is where we started.
Five things you can check before you sign anything.
Fine-tuning through OpenAI, Azure OpenAI or Google Vertex AI leaves the tuned model on their platform: you can call it, not download it. MediaAtlas hands over weights, adapters, datasets and evaluation sets, so you can run or move the model without us.
Hosted fine-tuning APIs need your data on their servers for training and on every call. Our models run on a laptop, a single GPU server or an air-gapped machine, and training happens in the EU under an NDA and data-processing agreement.
Evaluation €1,900, pilots from €3,900, programs from €20,000, retainers from €1,200 a month, all on the pricing page. Large consultancies such as Accenture or Deloitte quote per engagement, after a proposal cycle.
We freeze an evaluation set from your documents and score untuned models first, so every claim is a number on the same set. Self-serve platforms such as Together AI or Hugging Face AutoTrain run the training loop but leave the dataset and the evaluation to your team.
More than twenty open models and four datasets on Hugging Face, each with results: perplexity −51 % after continued pre-training on 1.78 billion Slovene tokens, speech recognition error 46.8 % → 22.3 %. Where we only reach parity with a public model, we say parity.
There is no account-management layer. The people who build your dataset and train your model are the people on the call, and the founder is involved in every training engagement.
We work best with teams that have a narrow, repeated language task and data that should not leave the house. These are the cases our process is designed around:
Founder · engineering lead
Tadej founded MediaAtlas in 2012 in Sevnica and today leads development. His work has moved from IoT platforms and NLP search over the COVID-19 research corpus to multi-agent AI assistants and, since 2026, a full Slovenian language, speech and vision model stack published on Hugging Face.
MediaAtlas started as a small software studio in Sevnica and grew into an AI research and development studio by following one rule: build the thing, then publish the evidence. That is why our models, datasets and results are public.
Today MediaAtlas is a small, remote-first engineering team headquartered in Sevnica. The engineers who build your dataset and train your model are the ones you work with directly.
MediaAtlas trains language models on a company's own data, builds the retrieval and document pipelines around them, and deploys the result on the customer's infrastructure. The customer receives the weights, datasets and evaluation sets. We also run two compliance SaaS products, Sensoram and EDC.
In Sevnica, Slovenia (Kajuhova 11, 8290 Sevnica). We are an EU company under EU law, and customer data stays inside the EU or inside the customer's own network.
Tadej Fius founded MediaAtlas in 2012 and leads the engineering. He is involved in every model-training engagement.
The price list is public. A one-day evaluation is €1,900, fixed-price pilots start at €3,900, full programs at €20,000 and monthly retainers at €1,200. Prices are in EUR, excluding VAT.
Yes. Weights, adapters, the datasets built for you, the evaluation sets and the provenance record are handed over. You can run, re-tune or move the model without us.
No. Training can run on EU capacity under an NDA and data-processing agreement, and the finished model runs on your own hardware, including air-gapped machines with no network. Nothing is required to leave the building.
The one-day evaluation tells you whether training is worth it at all. A measured pilot model follows within weeks, and every phase ends with a number on your own evaluation set.
No. Most projects run in the customer's language on a strong multilingual base. Slovenian is our public reference case, with more than twenty open models and published results.
Show us fifty real prompts and answers from your domain. Within a week we'll tell you what a model trained on your data could do, and what it would take.