# MediaAtlas

> MediaAtlas d.o.o. (Sevnica, Slovenia, founded 2012) fine-tunes language models on a company's own data and delivers the weights. Continued pre-training, supervised fine-tuning, dataset engineering and frozen-benchmark evaluation, trained on GPU capacity inside the EU. Also builds platforms for STEM education, verifiable data, hydrogen trading, cold-chain monitoring and working-time records.

## Flagship service: LLM fine-tuning and AI model training

- [AI Model Training (EN)](https://mediaatlas.si/ai-training.html): LLM fine-tuning as a service. Continued pre-training (CPT), supervised fine-tuning (SFT), dataset engineering, evaluation flywheel, GGUF/vLLM deployment, ASR and TTS fine-tuning. Fixed-price pilot, full program, monthly flywheel retainer. Customer owns weights, adapters, datasets and eval sets. Data stays in the EU; air-gapped training available.
- [Treniranje AI modelov (SL)](https://mediaatlas.si/sl/ai-training.html): Slovenian version of the same page.
- [Pricing (EN)](https://mediaatlas.si/ai-training-pricing.html) / [Cenik (SL)](https://mediaatlas.si/sl/ai-training-pricing.html): public price list. One-day evaluation on the customer's own documents EUR 1,900 (credited against a pilot); one-week trial under NDA; fixed-price pilots from EUR 3,900 (Pilot S) and EUR 12,000 (Pilot M); full programs from EUR 20,000; monthly flywheel retainers from EUR 1,200. EUR, excluding VAT.
- [Published models and datasets on Hugging Face](https://huggingface.co/MediaAtlas) (also under https://huggingface.co/texdata): 20+ open models (Slovenian LLMs 4B to 35B, biomedical research model, medical translation, Slovenian streaming ASR, Slovenian TTS) and public datasets, each with a model card and evaluation numbers.

How an engagement starts: (1) **Evaluation, one day, EUR 1,900** - the customer sends fifty anonymised examples and one task; we build a frozen set with gold answers, run up to five models (public Slovene models, open bases, the customer's candidate, a foreign API) through the same instruction, and return a table of measured results, verbatim outputs and a written recommendation on whether training is worth it. (2) One-week trial under NDA, model running on the customer's server, before-and-after numbers. (3) Pilot, program, retainer.

Typical use cases: customer support and ticket answering, document extraction and classification, internal knowledge assistants, agents with reliable tool calling, speech recognition and synthesis, language adaptation for under-served languages (Slovenian is the public reference implementation). Small fine-tuned models (4B to 12B) for narrow tasks, cascaded with a frontier fallback where needed.

Contact: info@mediaatlas.si, +386 40 42 33 99, https://cal.com/tfius-tfius

## Platforms

- [Mali Pokovci](https://mediaatlas.si/platforms/mali-pokovci.html): STEM education for children 6 to 14, LEGO robotics and visual programming.
- [SemantiCord](https://mediaatlas.si/platforms/semanticord.html): Memory Units, signed and hash-verified JSON envelopes with schema registry (RFC 8785 canonicalization).
- [EnergonX](https://mediaatlas.si/platforms/energonx.html): location-aware hydrogen trading marketplace with certification.
- [AssetManiac](https://mediaatlas.si/platforms/assetmaniac.html): multi-agent AI for blockchain and market intelligence.
- [Sensoram](https://mediaatlas.si/platforms/sensoram.html): cold-chain temperature and humidity monitoring, EN 12830 reports, hash-chained records.
- [EDC](https://mediaatlas.si/platforms/edc.html): Slovenian working-time records (ZEPDSV-A) with payroll export.
- [Motion Capture Studio](https://mediaatlas.si/platforms/motion-capture-studio.html): real-time 3D, motion capture, interactive media.

## Public evidence

- Evaluation method: frozen sets with gold answers, deterministic scoring where possible, an independent judge model for free text, verbatim outputs published alongside the numbers, and the weakest example always shown.
- Slovene language work is public: 20+ models on Hugging Face with evaluation results, a Slovene streaming ASR fine-tune (WER 46.8% to 22.3%), continued pre-training on 1.78 billion Slovene tokens (perplexity -51%), medical EN-SL machine translation, tool-calling models, and open Slovene datasets (medical SFT, medical eval, translation SFT, pretraining corpus).
- Morphology data built from Sloleks 3.0 (CC BY-SA 4.0) for Slovene declension, dual and imperative drills.

## Company

- [About](https://mediaatlas.si/about.html)
- [Team](https://mediaatlas.si/team.html)
- [Investors](https://mediaatlas.si/investors.html)
- [Contact](https://mediaatlas.si/contact.html)