Dias Kuatbekov

Machine Learning Engineer, Computer Vision Team, Biometrics

Kaspi.kz

The Leading FinTech Organization in Kazakhstan

Built production PAD and biometric vectorization models; reports 0% APCER at 0.01% BPCER and TAR 99.98% at FAR 1e-8.

• Engineered Presentation Attack Detection (PAD) models, achieving 0% APCER at 0.01% BPCER in production via custom architectures, targeted pre-training, and domain adaptation. • Applied SoTA methods for end-to-end training and inference pipelines for biometric vectorization models (TAR 99.98% at FAR 1e-8); Led R&D on explainability (xAI) techniques to for decision transparency and interpretability.

Computer Vision Research Scientist

Sergek Group

Leading AI-powered traffic & mobility platform

Researched and built knowledge-distillation pipelines for monocular depth estimation, reporting a 2.5x inference speedup on NVIDIA Jetson Orin.

• Conducted joint research on Knowledge Distillation (KD) for Monocular Depth Estimation (MDE), designing edge-optimized student architectures for traffic speed estimation in patrol cars. • Designed Knowledge Distillation pipelines for heteregenous teacher and student architectures (from transformers to CNNs) • Achieved competetive results with SOTA models like DepthAnythingV2 on KITTI and NYU benchmarks while achieving 2.5x inference speedup on NVIDIA Jetson Orin

Research Assistant

Nazarbayev University

Research university in Kazakhstan.

Applied GNNs and a VAE to molecular property prediction and generation; reports R2 = 0.96 for maximum light absorption prediction.

• Trained graph neural networks (GNN) to predict molecular photophysical properties, achieving an R2 = 0.96 for maximum light absorption prediction. • Trained a Variational Autoencoder (VAE) for de novo molecular generation, achieving, achieving a 91% synthesized novel molecules. • Boosted overall model performance and handled missing data by implementing using data imputation, custom loss functions, and advanced feature engineering using RDKit.

Research Assistant

Nazarbayev University

Research university in Kazakhstan.

Implemented and benchmarked FlashAttention in CUDA C, reporting a 300% speedup over a standard attention baseline.

• Implemented the FlashAttention algorithm in CUDA C, utilizing NVIDIA Nsight to benchmark performance and validate a 300% speedup over a standard attention baseline.

MSc, Machine Learning in Science

University of Nottingham

BSc, Computer Science

Nazarbayev University