The creator of DefaceID discusses why AI ought to clear up useful resource distribution issues and the way rising markets can compete globally.
In line with Gartner forecasts, agentic AI and post-quantum cryptography have entered the highest 10 strategic expertise traits for 2025. The blockchain expertise market will develop to $1.43 trillion by 2030 with a median annual development price of 90.1%, whereas the worldwide AI market will improve from $184 billion to $826 billion by 2030. It will require specialists who create programs that don’t simply course of knowledge, however act as brokers, making real-time selections, working with blockchain applied sciences in decentralized identification, and making certain most safety ranges. Kazakhstani professional Kenessary Koishybay has progressed from researcher to ML venture chief exactly in the course of the interval when these applied sciences transitioned from experimental stage to mass implementation, dealing with all typical challenges of scaling AI options – knowledge shortage, {hardware} limitations, and accuracy necessities.
In late 2024, the 28-year-old Chief Machine Studying Specialist at BTSDigital was awarded the Nationwide Business Prize “Technologies and Innovations” as AI Specialist of the 12 months for the DefaceID venture – a decentralized biometric authentication system on Web Laptop. He’s at the moment engaged on Digital ID system, which each resident of Kazakhstan now makes use of for authentication in authorities companies, developed the TargetAI digicam system for the nation’s largest corporations, participated in creating Business Prime Rewards for Fortune 500 corporations together with Apple and Tesla in Madrid, and likewise developed options for the automated visitors violation recognition system “Sergek”. He shared insights on the significance of decentralization over regulation in digital identification and biometric authentication programs, defined how Amazon’s method to pondering in ecosystems influenced him, and emphasised that AI shouldn’t be considered as magic.
“Architecture, data storage, verification – everything is built considering future connections of banks, services, government portals.”
You’ve progressed as a analysis assistant at Nazarbayev College with scientific publications on signal language to working at Amazon in Spain. How do AI approaches differ in these areas?
Working at Amazon, I confronted a scale that required fast adaptation. There, any technical answer is a part of a posh ecosystem the place all the pieces follows requirements: from logging to architectural compatibility. Every little thing is strictly documented, and each change is the results of interplay with a number of groups. This hardens you: you begin pondering not as an information scientist, however as an engineer in an ecosystem with tens of millions of customers.
Engaged on the Rebates program at Amazon – a part of Business Prime Rewards – I realized to assume by way of duty and fault tolerance. We developed options for calculating annual cashback for purchasers like Apple and Tesla – calculation errors may value tens of millions.
In Kazakhstan, a unique tempo. Selections are made sooner, groups are compact, a number of flexibility. This offers room for creativity, however usually lacks course of maturity: no clear MLOps requirements, mannequin model management. What’s thought of regular at Amazon – rollback pipeline or A/B monitoring – right here, it’s a must to construct from scratch. However when you do, you achieve a deeper understanding of the total ML lifecycle — not simply modeling, but additionally deployment, monitoring, and long-term upkeep. It turns you right into a builder, not only a person of instruments.
Business Prime Rewards is utilized by Apple, Google, Tesla, and Fortune 500 corporations. How did the expertise engaged on such a worldwide system affect your native initiatives like Digital ID?
This expertise laid the muse for my work on Digital ID. Though the venture is oriented towards the native market, we initially designed it with scaling and integration with authorities infrastructure in thoughts. Structure, knowledge storage, and verification are all constructed with future connections between banks, companies, and authorities portals in thoughts. With out understanding large-scale worldwide programs, this might have been unimaginable.
“DefaceID is not just technology, it’s a step toward personal digital autonomy”
Your DefaceID on the Web Laptop obtained recognition as one of the best AI venture of the yr. Blockchain for biometrics – is decentralization the longer term?
Decentralization isn’t just a development — it’s a structural response to the rising disaster of digital belief. In centralized programs, biometric knowledge is saved by third events, making a single level of failure. One leak, and the person completely loses management over their identification.
DeFaceID is essentially completely different: it’s a absolutely decentralized identification verification system. Each the logic and AI fashions — together with neural networks for liveness detection and face matching — are deployed immediately on-chain, on the Web Laptop. This ensures that no centralized get together can intercept or manipulate the verification course of.
Biometric templates are by no means saved or transmitted. As an alternative, customers confirm themselves via zero-knowledge proofs and on-chain mannequin execution, all inside trustless, tamper-proof canisters. This makes DeFaceID one of many first actually decentralized biometric programs — combining privateness, transparency, and cryptographic safety with out compromising scalability..
Can we are saying it is a Web3 various to Clear or Face ID? The US and EU actively regulate AI. How ought to decentralized identification and biometrics be thought of?
Sure, DefaceID is a Web3 various to centralized platforms. In contrast to them, our system doesn’t require belief in a single operator. Customers handle their very own biometrics and management identifiers, confirming identification via cryptographic protocols with out transmitting “raw” knowledge. This reduces dangers and makes the system extra GDPR-compliant.
Efficient AI regulation ought to mix authorized frameworks with technical safety mechanisms. AI ought to mix authorized frameworks with technical safety mechanisms. Decentralized architectures assist construct safety and privateness into the infrastructure itself, fairly than attempting to regulate all the pieces from above. That is particularly necessary now when applied sciences develop sooner than legal guidelines can adapt.
“Edge AI is a rethinking of AI system logic”
You optimized pc imaginative and prescient for edge units within the TargetAI venture, utilized by BI GROUP and Eurasian Sources Group. How do you stability accuracy and effectivity on restricted sources?
In Goal AI, we optimized face recognition and surveillance fashions for edge units — small packing containers with restricted reminiscence and compute. The important thing was to construct across the {hardware}, not simply the mannequin. We educated giant fashions for accuracy, then utilized quantization, pruning, and distillation to suit real-time constraints. Typically, buying and selling 2% accuracy gave us 5x sooner inference, which is essential when selections are made on the fly.
Engaged on “Sergek” – the visitors violation processing system – how did you combat false positives?
In Sergek, the nationwide visitors violation system, false positives have been unacceptable. Each AI-generated occasion was verified by a human, and customers may enchantment any nice. However edge challenges remained — low gentle, climate modifications, occlusions. We constantly up to date coaching datasets with edge circumstances and used context-aware fashions that took into consideration elements like time of day and visibility. Right here, accuracy wasn’t nearly numbers — it was about public belief.
Edge AI isn’t simply “AI that runs on devices.” It’s a brand new philosophy: make sensible programs work below real-world constraints — offline, below stress, with minimal sources, however with out compromising on equity, pace, or safety.
“Modern systems lack deep protection against falsification”
As Chief Machine Studying Specialist at BTSDigital, a serious contributor to Digital ID utilized by each Kazakhstani, what do you assume fashionable digital identification options lack?
Deep safety in opposition to falsification, particularly in biometrics. Many options solely test for “face presence,” not recognizing fakes. Distributors don’t present resistance to 3D masks, deepfakes.
In Digital ID, we take note of liveness detection. We use neural community fashions educated on actual assaults, combining indicators from the digicam and person habits. This identifies subtle spoofing makes an attempt that cross by customary options and given excessive safety necessities, we’re proud to say our system is among the greatest on the planet.
Along with your expertise, from 6 scientific publications to working with Fortune 500, what undervalued expertise do ML engineers want at present?
Working with small and noisy knowledge is an actual artwork, undervalued in opposition to the backdrop of the “more data – better results” development. In actual initiatives, you not often have an excellent million-sample dataset. The power to extract that means from restricted info turns into important.
System pondering expertise are additionally necessary: information of CI/CD, MLOps infrastructure, and mannequin monitoring. A mannequin just isn’t a .pt file, however a part of a posh ecosystem, and the power to correctly combine it’s an undervalued however beneficial ability.
“It doesn’t matter where you are – what matters is your contribution.”
Kazakhstan isn’t probably the most conventional AI hub, however your initiatives are utilized by tens of millions of individuals and main companies. How can specialists from rising markets compete globally?
The market is turning into decentralized, creating alternatives for specialists from growing international locations. The principle factor is to not be afraid to transcend the native market. Participation in open-source, publications, exercise on GitHub, and Kaggle – methods to make your self recognized no matter geography.
I progressed from native startups in Kazakhstan to Amazon in Spain due to my technical background, English information, and drive to develop. It doesn’t matter the place you’re – what issues is your contribution: code, concepts, analysis, options.
What world downside would you clear up utilizing AI?
I’d need AI to assist optimize world useful resource distribution – meals, vitality, and medication. The world has sufficient expertise to reduce poverty and starvation, however lacks coordination and transparency. AI can think about tens of millions of parameters and discover optimum supply and redistribution paths.
I’d additionally rethink trade attitudes. For too lengthy, it lived below the slogan “more data is better,” forgetting that high quality is extra necessary than quantity. An excessive amount of consideration is paid to new architectures and little to mannequin stability. AI is perceived as magic, not engineering. The trade ought to assume extra by way of systemic duty: testability, reproducibility, logging, fallback mechanisms, A/B testing, and mannequin resolution auditing. As a result of a mannequin in isolation is only a prototype, and actual worth begins when it’s reliably built-in right into a workflow and doesn’t fail below advanced situations.
