Director of Digital Innovation at Diamdor
Kazakhstan
Aleksandr Polezhaev
Council Member
Aleksandr Polezhaev is an independent researcher and technical author with expertise spanning cloud architecture, production AI systems, and distributed systems engineering, currently based at the intersection of academic research and applied industry work. With a research and professional career extending from 2020 to the present, he has built a body of work that bridges rigorous systems engineering with practical deployment realities, contributing original frameworks through peer-reviewed publications, technical monographs, and pre-registration inventions.
Since 2021, Aleksandr Polezhaev has served as Director of Digital Innovation at Diamdor, where he leads the design and deployment of AI-integrated software systems across cloud and distributed environments. In this role, he is responsible for technical strategy, architecture decisions, and engineering innovation initiatives, translating research-grade concepts into production-ready systems at scale.
Alongside his work at Diamdor, Aleksandr Polezhaev has maintained an independent research practice since 2020, producing original engineering frameworks across three interconnected systems. The Adaptive AI Resource Allocation Engine (AARAE) was evaluated against a threshold-based autoscaler baseline across 50 independent runs on production-derived multi-tenant cloud workloads. The system achieved reallocation command generation latency of 87 ms mean and 143 ms at p95 on problem instances up to 10,000 workload units and 500 physical nodes, delivering up to a 40% throughput improvement under bursty load conditions and a 20–35% reduction in resource waste across workloads with variable per-request resource intensity. At larger scale - 50,000 workload units and 2,000 nodes - mean latency reached 610 ms while maintaining solution quality within 5-8% of the integer programming optimum. The architecture integrates a User Behavioral Profiling Module producing 128-dimensional behavioral signature vectors, a dual-stage Temporal Predictive Load Estimator combining GRU and Transformer components through a learned fusion layer, a Constraint-Aware Resource Redistribution Controller using pre-computed sparse constraint matrices, and a Closed-Loop Calibration Subsystem with elastic weight consolidation to prevent catastrophic forgetting during online adaptation.
His Secure Behavioral Authentication Framework addressed continuous identity verification in distributed environments without reliance on password-based mechanisms. Evaluated across 64 users with sessions incorporating adversarially generated mimicry samples from trained human mimics, the multi-modal framework achieved an Equal Error Rate of 1.02% — a 70.9% improvement over the keystroke-only baseline of 3.50% — while reducing mimicry False Acceptance Rate from 11.2% to 0.003%. Identity substitution detection latency improved to 31.4 seconds, 35% faster than the single-modality baseline, and session suspension false positive rate fell from 4.1% to 1.6%. The framework fuses five concurrent behavioral signal channels through a two-tier stacked ensemble classifier with session-level LSTM temporal encoding, per-user threshold calibration, and adaptive profile maintenance with adversarial drift detection.
Since 2021, Aleksandr Polezhaev has served as Director of Digital Innovation at Diamdor, where he leads the design and deployment of AI-integrated software systems across cloud and distributed environments. In this role, he is responsible for technical strategy, architecture decisions, and engineering innovation initiatives, translating research-grade concepts into production-ready systems at scale.
Alongside his work at Diamdor, Aleksandr Polezhaev has maintained an independent research practice since 2020, producing original engineering frameworks across three interconnected systems. The Adaptive AI Resource Allocation Engine (AARAE) was evaluated against a threshold-based autoscaler baseline across 50 independent runs on production-derived multi-tenant cloud workloads. The system achieved reallocation command generation latency of 87 ms mean and 143 ms at p95 on problem instances up to 10,000 workload units and 500 physical nodes, delivering up to a 40% throughput improvement under bursty load conditions and a 20–35% reduction in resource waste across workloads with variable per-request resource intensity. At larger scale - 50,000 workload units and 2,000 nodes - mean latency reached 610 ms while maintaining solution quality within 5-8% of the integer programming optimum. The architecture integrates a User Behavioral Profiling Module producing 128-dimensional behavioral signature vectors, a dual-stage Temporal Predictive Load Estimator combining GRU and Transformer components through a learned fusion layer, a Constraint-Aware Resource Redistribution Controller using pre-computed sparse constraint matrices, and a Closed-Loop Calibration Subsystem with elastic weight consolidation to prevent catastrophic forgetting during online adaptation.
His Secure Behavioral Authentication Framework addressed continuous identity verification in distributed environments without reliance on password-based mechanisms. Evaluated across 64 users with sessions incorporating adversarially generated mimicry samples from trained human mimics, the multi-modal framework achieved an Equal Error Rate of 1.02% — a 70.9% improvement over the keystroke-only baseline of 3.50% — while reducing mimicry False Acceptance Rate from 11.2% to 0.003%. Identity substitution detection latency improved to 31.4 seconds, 35% faster than the single-modality baseline, and session suspension false positive rate fell from 4.1% to 1.6%. The framework fuses five concurrent behavioral signal channels through a two-tier stacked ensemble classifier with session-level LSTM temporal encoding, per-user threshold calibration, and adaptive profile maintenance with adversarial drift detection.
The third system, the Adaptive Consistency Orchestration Protocol (ACOP), was evaluated on a 16-node cluster running six domain microservices across dedicated PostgreSQL 15.2 instances with Apache Kafka 3.6 as message transport. ACOP achieved 96.1% of the eventually consistent baseline throughput at a mean latency of 4.2 ms under nominal load, and 94.3% at p99 7.1 ms under high-concurrency conditions of 12,000 TPS - a load under which Two-Phase Commit failed entirely. In a network-degraded scenario with 150 ms added round-trip time and 0.5% packet loss, ACOP maintained 91.7% of baseline throughput. It delivered a 9.9× throughput advantage over Two-Phase Commit at nominal load, a 35–50% commit latency reduction compared to leader-based consensus protocols, and a 10.8× reduction in blocking cross-service reads versus pessimistic locking, supported by 91.4% Markov prediction accuracy over a continuous 72-hour run.
Aleksandr Polezhaev has published three technical monographs through Lambert Academic Publishing. "Artificial Intelligence in Production: Practical Engineering Beyond Theory" addresses the engineering challenges of deploying and operating AI systems in real-world environments, shifting emphasis from model-centric thinking toward system-oriented design and offering frameworks for building resilient, maintainable, and cost-effective industrial AI systems. "Architecting Scalable Cloud Systems: From Startup to Enterprise" presents a practical roadmap for designing and evolving cloud-native architectures across organizational growth stages, covering service decomposition, containerization and orchestration, auto-scaling strategies, serverless computing, multi-cloud patterns, and adaptive AI-driven resource allocation. "Secure Software Design: Engineering Applications for the Post-Password Era" examines the transition from password-based authentication to cryptographic identity systems in distributed environments, analyzing zero-trust architecture as a design constraint and providing a systems-level framework for authentication mechanisms that function reliably under real-world distributed conditions.
His peer-reviewed journal and conference work includes "Security as a Feature: The Next Competitive Advantage in Software Products," published in the Austrian Journal of Technical and Natural Sciences; "Why AI Projects Fail in Production and How Engineers Should Build Them," presented at the CVII International Scientific and Practical Conference through Internauka; and "Event-Driven Architectures Are Replacing Monoliths: The Future of Software Design," published in Universum: Technical Sciences.
Aleksandr Polezhaev has published three technical monographs through Lambert Academic Publishing. "Artificial Intelligence in Production: Practical Engineering Beyond Theory" addresses the engineering challenges of deploying and operating AI systems in real-world environments, shifting emphasis from model-centric thinking toward system-oriented design and offering frameworks for building resilient, maintainable, and cost-effective industrial AI systems. "Architecting Scalable Cloud Systems: From Startup to Enterprise" presents a practical roadmap for designing and evolving cloud-native architectures across organizational growth stages, covering service decomposition, containerization and orchestration, auto-scaling strategies, serverless computing, multi-cloud patterns, and adaptive AI-driven resource allocation. "Secure Software Design: Engineering Applications for the Post-Password Era" examines the transition from password-based authentication to cryptographic identity systems in distributed environments, analyzing zero-trust architecture as a design constraint and providing a systems-level framework for authentication mechanisms that function reliably under real-world distributed conditions.
His peer-reviewed journal and conference work includes "Security as a Feature: The Next Competitive Advantage in Software Products," published in the Austrian Journal of Technical and Natural Sciences; "Why AI Projects Fail in Production and How Engineers Should Build Them," presented at the CVII International Scientific and Practical Conference through Internauka; and "Event-Driven Architectures Are Replacing Monoliths: The Future of Software Design," published in Universum: Technical Sciences.