AkylAI — the first AI assistant in Kyrgyz language

The founders of the startup AkylAI told how they created the first AI assistant in the Kyrgyz language. Read more in our article.

Timur Turatali, Nursultan Bakashov, Ulan Abdurazakov, Bishkek, founders of AkylAI

How it all started

We started working together in 2022. AkylAI has three founders: Ulan Abdurazakov, Timur Turatali and Nursultan Bakashov. In the beginning, we just ran our own courses on Data Science and machine learning. We also founded the AI-community of Kyrgyzstan. This is a community of specialists engaged in artificial intelligence, around which we organized different courses.

Last year, we had the idea that the digital world was evolving, and with the arrival of ChatGPT, everything changed dramatically. At that time, ChatGPT did not know the Kyrgyz language, and we decided to build a language corpus — a collection of words and texts to train the model.

We started the corpus in May 2023, and within three months we had collected over 300 million Kyrgyz words. We were assisted by partners including the High Technology Park of the Kyrgyz Republic, the Kyrgyz-Turkish Manas University and the Kyrgyz State Technical University.

Team

After building the first corpus, we decided to create an LLM in Kyrgyz based on Llama. In August 2023, we released the first beta version of AkylAI. It was a website with a chatbot that answered queries in Kyrgyz. Initially, we thought we would stop there.

We thought for a long time how to develop the project further. We noticed that Alice's columns had become popular and decided to create our own column. In this we were strongly supported by the High Tech Park, which offered to complete the AkylAI project and then package it into a device. ChatGPT already knows Kyrgyz and solves many issues. But we liked the idea of creating a device for the home that you can communicate with.

We started working on the hardware part in January 2024. First we gathered a team. The first to find an industrial designer was Oleg Steel Drake Yarmakov, who started doing design and prototyping. We also found an electronics engineer, Egor Garayushchenko, who can assemble the board for the device.

One of the most difficult tasks is to make the model speak Kyrgyz. This language was not initially supported, so we started from scratch. To train the model we recorded more than 24 hours of text with the actress in the studio. At first the voice sounded with an accent, but by April we were able to achieve the right quality.

Target audience

We have a web version with 15 000 users. It is free and allows you to communicate and solve tasks in Kyrgyz. In the future, we see AkylAI as not just an assistant for the home, but a real companion and friend who is always ready to listen and give advice. Something like a psychologist or a therapist.

We made the voice available in open source to support the community and programmers. Now anyone who wants can take this voice and code and use it for their own purposes.

In June, we presented the first five prototypes that can speak and answer questions. There is still a lot of work to be done, but the first step is important for us — even though nothing was unclear or difficult, we managed.

Incubation programs

May 2024 walked into the US Launch gas pedal from Jason Calacanis' Founder University. Here presented a project with a focus on mental health and AI Buddy. We are going through the accelerator and preparing for a demo day in August.

Last year we participated in the Unicorn from KG program from the High Tech Park, which provided us with seed funding. As part of this program, we received one month of training from Draper University.

The High Tech Park believed in us even more than we believed in ourselves. They supported us not only financially, but also morally. We would also like to thank the European Bank for Reconstruction and Development, EBRD. We participated in their Star Venture program and became winners. We thank all partners who helped us and believed in our project. This is not the work of one person, but the work of our entire team.

Challenges

Training an AI model is always a confusing endeavor. You do not know what the result will be, and that means stress. That was the case with Voice. We trained it three or four times to reach a good level.

The second problem is the modular part, the electronics. We manufactured everything at our place, but we ordered from China. We had a problem with a module from Israel, where there were difficulties with logistics. We had to reorder through China.

The problems were with Hardware - loading the model into the speaker so that they worked. We had enough manpower to assemble the first model, but now we need more experienced and skilled people.

To understand the scale: Amazon spent $20 billion to $43 billion on its Alexa, according to various estimates, and Alice cost about $20 million, according to various estimates. Much of that cost probably went to human capital.

The key skill of a startup is the ability to live in uncertainty, when nothing is certain, you don't know if the product will turn out or not. Learn to accept and enjoy this state. It is important not to be afraid of uncertainty and to find opportunities in it. The hardest thing is to take the first step.

Achievements

We consider our victory to be the creation of the first Kyrgyz voice assistant. This is a great achievement not only for us, but for the whole country. The project was bold, interesting and challenging. Not everyone would dare to do such a thing.

We have been accelerated by Launh. It is cool that we, people from Central Asia, can create cool ideas for the whole world even being here. The main thing is to remove the frames in your head.

Plans

We plan to enter the US market by the end of this year. This is important because we need investment and financial resources. There are few people in Central Asia who are ready to provide the necessary amount of funds, and the market here is much smaller. There are more opportunities in the US.

We have already presented our American product and made an MVP, which is being tested by a focus group. Our goal is to attract more users and refine the column in parallel. By the end of the year, we will finalize the speaker and make it a product ready for production in factories.

I wish startups to be brave, to go forward, to try. Fail and fail again until you succeed.