An Azerbaijani LLM on SageMaker: how Azercell tackled a low-resource language

Azerbaijani telecom operator Azercell has built its own language model on Amazon SageMaker AI. The company set itself an ambitious goal: a production-grade LLM for a morphologically complex language with little ready-made data and no existing solutions on the market.

Why Azerbaijani is a hard case

Azerbaijani is a typical agglutinative language with rich morphology. A single word can carry many suffixes, each of which can radically change its meaning and grammatical function. That calls for very different approaches to tokenization and model training than Indo-European languages do.

On top of the morphology comes a critical constraint: there is far less open training data in Azerbaijani than in English, Russian, or Spanish. Standard LLM training methods, refined on large text corpora, don't transfer directly.

• Morphological complexity requires specialized tokenization
• Data scarcity: 100+ times less text than for major languages
• No existing examples or best practices for Azerbaijani LLMs
• Foundation models trained on English-language data need to be adapted
• The model has to integrate with production telecom systems

How Azercell solved it

The company partnered with the AWS Generative AI Innovation Center. Over six intensive weeks, specialists from both sides built a production-ready framework on Amazon SageMaker.

The solution had several key components: careful preparation and cleaning of the available data, specialized tokenization that accounts for Azerbaijani morphology, and a training process optimized for smaller datasets. Rather than training from scratch on an Azerbaijani corpus, the engineers used transfer learning to adapt already-trained models.

The result: two roles for the model

Azercell's model serves two purposes. First, it runs a customer-facing chatbot that helps subscribers with questions about services and plans in Azerbaijani.

Second, the model supports internal business processes: handling incoming requests, analyzing call-center speech, classifying issues, and generating personalized service recommendations. Focusing on Azerbaijani avoids meaning lost in translation and ensures the model understands local context and nuances of speech.

What it means

This is the first public example of a fully functional Azerbaijani LLM developed on cloud infrastructure. The case shows that cloud platforms can adapt LLMs not only to low-resource languages but also to specific industry tasks.

For other companies in the region, the message is clear: investing in a proprietary language model is realistic and achievable within a few weeks.