The creator of DefaceID discusses why AI should solve resource distribution problems and how emerging markets can compete globally.
According to Gartner forecasts, agentic AI and post-quantum cryptography have entered the top 10 strategic technology trends for 2025. The blockchain technology market will grow to $1.43 trillion by 2030 with an average annual growth rate of 90.1%, while the global AI market will increase from $184 billion to $826 billion by 2030. This will require specialists who create systems that don’t just process data, but act as agents, making real-time decisions, working with blockchain technologies in decentralized identification, and ensuring maximum protection levels. Kazakhstani expert Kenessary Koishybay has progressed from researcher to ML project leader precisely during the period when these technologies transitioned from experimental stage to mass implementation, facing all typical challenges of scaling AI solutions – data scarcity, hardware limitations, and accuracy requirements.
In late 2024, the 28-year-old Chief Machine Learning Specialist at BTSDigital was awarded the National Business Prize “Technologies and Innovations” as AI Specialist of the Year for the DefaceID project – a decentralized biometric authentication system on Internet Computer. He is currently working on Digital ID system, which every resident of Kazakhstan now uses for authentication in government services, developed the TargetAI camera system for the country’s largest companies, participated in creating Business Prime Rewards for Fortune 500 companies including Apple and Tesla in Madrid, and also developed solutions for the automated traffic violation recognition system “Sergek”. He shared insights on the importance of decentralization over regulation in digital identity and biometric authentication systems, explained how Amazon’s approach to thinking in ecosystems influenced him, and emphasized that AI shouldn’t be viewed as magic.
“Architecture, data storage, verification – everything is built considering future connections of banks, services, government portals.”
You’ve progressed as a research assistant at Nazarbayev University with scientific publications on sign language to working at Amazon in Spain. How do AI approaches differ in these regions?
Working at Amazon, I faced a scale that required quick adaptation. There, any technical solution is part of a complex ecosystem where everything follows standards: from logging to architectural compatibility. Everything is strictly documented, and every change is the result of interaction with several teams. This hardens you: you start thinking not as a data scientist, but as an engineer in an ecosystem with millions of users.
Working on the Rebates program at Amazon – part of Business Prime Rewards – I learned to think in terms of responsibility and fault tolerance. We developed solutions for calculating annual cashback for clients like Apple and Tesla – calculation errors could cost millions.
In Kazakhstan, a different pace. Decisions are made faster, teams are compact, lots of flexibility. This gives room for creativity, but often lacks process maturity: no clear MLOps standards, model version control. What’s considered normal at Amazon – rollback pipeline or A/B monitoring – here, you have to build from scratch. But once you do, you gain a deeper understanding of the full ML lifecycle — not just modeling, but also deployment, monitoring, and long-term maintenance. It turns you into a builder, not just a user of tools.
Business Prime Rewards is used by Apple, Google, Tesla, and Fortune 500 companies. How did the experience working on such a global system influence your local projects like Digital ID?
This experience laid the foundation for my work on Digital ID. Although the project is oriented toward the local market, we initially designed it with scaling and integration with government infrastructure in mind. Architecture, data storage, and verification are all built with future connections between banks, services, and government portals in mind. Without understanding large-scale international systems, this would have been impossible.
“DefaceID is not just technology, it’s a step toward personal digital autonomy”
Your DefaceID on the Internet Computer received recognition as the best AI project of the year. Blockchain for biometrics – is decentralization the future?
Decentralization is not just a trend — it’s a structural response to the growing crisis of digital trust. In centralized systems, biometric data is stored by third parties, creating a single point of failure. One leak, and the user permanently loses control over their identity.
DeFaceID is fundamentally different: it is a fully decentralized identity verification system. Both the logic and AI models — including neural networks for liveness detection and face matching — are deployed directly on-chain, on the Internet Computer. This ensures that no centralized party can intercept or manipulate the verification process.
Biometric templates are never stored or transmitted. Instead, users verify themselves through zero-knowledge proofs and on-chain model execution, all within trustless, tamper-proof canisters. This makes DeFaceID one of the first truly decentralized biometric systems — combining privacy, transparency, and cryptographic security without compromising scalability..
Can we say this is a Web3 alternative to Clear or Face ID? The US and EU actively regulate AI. How should decentralized identification and biometrics be considered?
Yes, DefaceID is a Web3 alternative to centralized platforms. Unlike them, our system doesn’t require trust in one operator. Users manage their own biometrics and control identifiers, confirming identity through cryptographic protocols without transmitting “raw” data. This reduces risks and makes the system more GDPR-compliant.
Effective AI regulation should combine legal frameworks with technical protection mechanisms. AI should combine legal frameworks with technical protection mechanisms. Decentralized architectures help build security and privacy into the infrastructure itself, rather than trying to control everything from above. This is especially important now when technologies develop faster than laws can adapt.
“Edge AI is a rethinking of AI system logic”
You optimized computer vision for edge devices in the TargetAI project, used by BI GROUP and Eurasian Resources Group. How do you balance accuracy and efficiency on limited resources?
In Target AI, we optimized face recognition and surveillance models for edge devices — small boxes with limited memory and compute. The key was to build around the hardware, not just the model. We trained large models for accuracy, then applied quantization, pruning, and distillation to fit real-time constraints. Sometimes, trading 2% accuracy gave us 5x faster inference, which is crucial when decisions are made on the fly.
Working on “Sergek” – the traffic violation processing system – how did you fight false positives?
In Sergek, the national traffic violation system, false positives were unacceptable. Every AI-generated event was verified by a human, and users could appeal any fine. But edge challenges remained — low light, weather changes, occlusions. We continuously updated training datasets with edge cases and used context-aware models that took into account factors like time of day and visibility. Here, accuracy wasn’t just about numbers — it was about public trust.
Edge AI isn’t just “AI that runs on devices.” It’s a new philosophy: make smart systems work under real-world constraints — offline, under stress, with minimal resources, but without compromising on fairness, speed, or security.
“Modern systems lack deep protection against falsification”
As Chief Machine Learning Specialist at BTSDigital, a major contributor to Digital ID used by every Kazakhstani, what do you think modern digital identification solutions lack?
Deep protection against falsification, especially in biometrics. Many solutions only check for “face presence,” not recognizing fakes. Vendors don’t provide resistance to 3D masks, deepfakes.
In Digital ID, we pay attention to liveness detection. We use neural network models trained on real attacks, combining signals from the camera and user behavior. This identifies sophisticated spoofing attempts that pass by standard solutions and given high security requirements, we are proud to say our system is one of the best in the world.
With your experience, from 6 scientific publications to working with Fortune 500, what undervalued skills do ML engineers need today?
Working with small and noisy data is a real art, undervalued against the backdrop of the “more data – better results” trend. In real projects, you rarely have an ideal million-sample dataset. The ability to extract meaning from limited information becomes critical.
System thinking skills are also important: knowledge of CI/CD, MLOps infrastructure, and model monitoring. A model is not a .pt file, but part of a complex ecosystem, and the ability to properly integrate it is an undervalued but valuable skill.
“It doesn’t matter where you are – what matters is your contribution.”
Kazakhstan isn’t the most traditional AI hub, but your projects are used by millions of people and major corporations. How can specialists from emerging markets compete globally?
The market is becoming decentralized, creating opportunities for specialists from developing countries. The main thing is not to be afraid to go beyond the local market. Participation in open-source, publications, activity on GitHub, and Kaggle – ways to make yourself known regardless of geography.
I progressed from local startups in Kazakhstan to Amazon in Spain thanks to my technical background, English knowledge, and drive to develop. It doesn’t matter where you are – what matters is your contribution: code, ideas, research, solutions.
What global problem would you solve using AI?
I would want AI to help optimize global resource distribution – food, energy, and medicine. The world has enough technology to minimize poverty and hunger, but lacks coordination and transparency. AI can consider millions of parameters and find optimal delivery and redistribution paths.
I would also reconsider industry attitudes. For too long, it lived under the slogan “more data is better,” forgetting that quality is more important than volume. Too much attention is paid to new architectures and little to model stability. AI is perceived as magic, not engineering. The industry should think more in terms of systemic responsibility: testability, reproducibility, logging, fallback mechanisms, A/B testing, and model decision auditing. Because a model in isolation is just a prototype, and real value begins when it’s reliably integrated into a workflow and doesn’t fail under complex conditions.
