Language is integral to artificial intelligence. But most AI applications are trained on the English language, which risks creating biases and excluding entire cultures from the digital wave. A team of young developers in the Kyrgyz Republic is working to ensure their native language isn't left behind. Their creation, AkylAI, represents the world's first artificial intelligence system fully compatible with the Kyrgyz language. AkylAI is also one of the EBRD’s Star Venture start-up beneficiaries which the Bank is helping to refine its strategy and expand its business abroad.
"When we told people outside Kyrgyz Republic that we were building a language model in our native tongue, they were confused," recalls Timur, one of AkylAI's co-founders. " 'Why would your country need AI in its own language?' they asked. "We explained that while Russian remains predominantly used in our country, we're working to preserve and promote our native Kyrgyz language."
How to build a large language model
AkylAI’s story began five years ago when Nursultan and Ulan collaborated on a tech bootcamp, teaching students data science and machine learning. Their initiative laid the foundation for Kyrgy Republic's first AI community. The team expanded when Timur returned from Europe, bringing fresh perspectives to their venture.
As ChatGPT gained prominence on a global scale, the team identified a crucial gap: the platform's incompatibility with the Kyrgyz Republic. They started an ambitious project to build a comprehensive language corpus - collecting over 300 million Kyrgyz words within just three months. Using this data, they fine-tuned an open-source language model based on Llama, leading to the 2023 release of their beta version: a web-based chatbot responding entirely in Kyrgyz.
The AkylAI virtual assistant
Their next milestone was developing a hardware device capable of voice-to-voice communication in Kyrgyz. According to Timur, this represented their biggest challenge, requiring extensive work on voice recognition and synthesis, including over 24 hours of studio recordings with an actress to train the model.
"When you work in a startup, there's a lot of uncertainty," Timur notes. "You have to learn to accept and enjoy this state. And just keep going."
With support from the EBRD's Star Venture programme – supported by the Small Business Impact Fund* and guidance from Cambridge University experts – the team refined their business strategy. "The support from the EBRD and Cambridge really helped us create a system and structure around how to think about our business. This is kind of the bigger picture, and you are here. These are your strengths and weaknesses."
Today, AkylAI's web platform serves 15,000 users, actively contributing to the digital presence of the Kyrgyz language. The team has successfully developed five prototypes capable of speaking and answering questions, although they are taking a measured approach to commercialisation.
The EBRD’s Small Business Impact Fund (Ireland, Italy, Japan, Korea, Luxembourg, Norway, Sweden, Switzerland, TaiwanBusiness– EBRD Technical Cooperation Fund and the USA).