How do you create an AI for a language that is virtually non-existent in modern neural networks? Bakhtiyar Mamedov started working with artificial intelligence back in 10th grade, and today he is developing TakykAI—the first AI startup for the Turkmen language. Discover his journey, the development of custom AI models, and his plans to expand into the Central Asian market in this feature by The Tech (a Kazakhstani portal covering startups and venture investments in Central Asia and the Caucasus), prepared in collaboration with the Startup Academy of Turkmenistan.
Bakhtiyar Mammedov, Kuala Lumpur, AI Developer at TakykAI
How It All Started
I started working with artificial intelligence back in the 10th grade, beginning with recommendation systems—the algorithms that suggest what users might like, such as on YouTube or e-commerce marketplaces.
By the time I graduated from high school, I had already built a solid knowledge base and completed several successful projects on fine-tuning large language models—adapting existing neural networks for specific tasks. However, these types of freelance assignments became quite boring for me, and I started thinking about what would truly keep me engaged over the long term. That was when I decided to focus on advancing the Turkmen language in the field of AI.
At that point, there were virtually no neural networks that understood Turkmen—neither text-based nor voice-based, meaning systems that could perform text-to-speech or speech recognition. I started from square one—by gathering data. This project has been my main focus for two years now.
Initially, I invested not only a huge amount of time into training the models, but also all of my spare money. Training or running any neural network required renting expensive compute resources—high-performance GPUs used for processing models. Over time, as promising results began to show, people took notice and started offering partnerships. Out of several offers, I chose the most compelling one for the future of the project.
With the onboarding of partners, our growth accelerated dramatically. We were able to assemble a dedicated team of developers, and we finally officially launched.
The Biggest Challenge Was the Data
The main hurdle was not building the AI itself, but gathering the data. Information in Turkmen was extremely scarce and scattered all across the web. I had to write numerous web-scraping scripts to collect data automatically and experiment with different architectural approaches to model building.
Speech models proved to be the most difficult. There were no ready-made solutions—free or paid—capable of recognizing or synthesizing Turkmen speech, while commercial recording studios charged steep fees to record hours of required audio material.
Through crowdsourcing—getting a large crowd of different people to contribute small amounts of work—we managed to train a speech recognition model. Initially, it operated only on fragmented text and audio datasets and had an error rate of about 10%.
Since then, we have successfully brought that error rate down to 3%. Building directly on top of this model, we were then able to gather the necessary data and train our first text-to-speech model, which converts written text into natural-sounding voice.
What We Have Built
At present, we have developed some of the world’s top-performing Turkmen language models, a speech synthesis model, and a speech recognition model. This also stands as the first AI startup dedicated to the Turkmen language.
Contrary to initial expectations, the primary target audience turned out to be small and medium-sized businesses, followed by end users of our consumer app, which functions similarly to ChatGPT.
We acquired our first users through niche Telegram channels and targeted search advertising.
How AI Changed the Workflow
Currently, AI is the core tool in my development process. It accelerates hypothesis testing and the creation of working demos down to a few days or even hours. Just a year ago, I could not delegate full-scale development to a neural network and coded all critical modules and security architecture manually. However, over the past year, AI has advanced to the point where I can hardly write better code than it does.
That said, I still manually review all AI-generated code.
Future Plans
By the end of the year, we plan to complete integrations with our B2B clients, streamline and accelerate the onboarding process for corporate users, and roll out paid subscription tiers within our consumer application.
Looking further ahead, I am setting my sights on the broader Central Asian market. We are currently launching market research to evaluate the demand across different countries for sovereign, national AI solutions hosted on their own domestic infrastructure.
The thought of giving up on the project has never crossed my mind. What keeps driving me forward is the vision itself and the tangible results we have already achieved. /// The Tech, 18 August 2026