About
I'm a deep learning researcher working across computer vision, multimodal learning, and scientific machine learning. My work runs from wearable-sensor and video activity recognition to gravitational-wave astrophysics, climate modeling, and medical imaging — held together by a consistent emphasis on interpretability and robustness under distribution shift.
I completed my MSc in Data Science at Nazarbayev University (Astana, Kazakhstan) and my BSc in Computer Systems Engineering at UET Peshawar. My research has taken me across Pakistan, Taiwan, and Kazakhstan — spanning multimodal activity recognition, scientific ML for astrophysics and climate, computer vision under adverse conditions, biomedical imaging, and ML on resource-constrained hardware.
I'm actively seeking PhD research opportunities in deep learning for sensing, perception, and scientific applications. Alongside research, I teach English independently in Astana.
Research interests
Computer Vision
Video and image understanding — from human activity recognition to crowd counting under adverse weather and medical image segmentation, with attention to where models break under real-world conditions.
Multimodal Learning
Combining complementary streams — video, wearable inertial sensors, and more — through learned, input-conditioned fusion that outperforms fixed rules while staying interpretable.
Signal & Time-Series ML
Learning from sequential sensor and physiological data with architectures — CNN-LSTM hybrids, 1D/3D CNNs — that respect temporal structure and channel-level physics.
Scientific Machine Learning
ML for physical systems — classifying nuclear equations of state from gravitational-wave signals, and variance-preserving deep models for climate bias-correction.
Interpretable & Robust DL
Grad-CAM, integrated gradients, channel-importance, and leave-one-subject-out protocols to ensure models generalize under distribution shift and their decisions can be audited.
Efficient & On-Device ML
Training and benchmarking models under tight memory and compute limits — architecture-efficiency studies and ablations to find what matters most on constrained hardware.
Selected projects
Learned Fusion for Activity Recognition 84.9% Detection F1
Multimodal human activity recognition on the C-MHAD benchmark, fusing wearable IMU and RGB video. An input-conditioned gated-fusion head beat maximum-score fusion by +3.8% (9/10 seeds) under a strict leave-one-subject-out protocol, with stable, interpretable per-class trust weights.
XEOS-NET — Nuclear EOS from Gravitational Waves >95% accuracy
A dual-pipeline CNN classifying the nuclear equation of state from core-collapse supernova gravitational-wave signals — raw 1D waveforms and 2D wavelet scalograms — with Grad-CAM, integrated gradients, and temporal-masking ablation for physical interpretability.
CMIP6 Climate Bias-Correction Scientific ML
A Conv1D+LSTM model with a custom variance-preserving loss for correcting systematic temperature bias in climate model projections, substantially reducing both mean bias and RMSE. Validated with Mann-Kendall trend and seasonal analysis to confirm the correction preserved physical structure rather than flattening variability.
Transfer Learning for Tumor Segmentation 0.719 Dice
Compared single-domain vs. multi-domain pre-training for liver-tumor segmentation with nnU-Net v2, showing that anatomical diversity beats task-specific depth in transfer learning — an 8% Dice gain over baseline.
Rain-Robust Crowd Counting NTUT · Taiwan
A crowd-counting model robust to rain, trained on synthetic adverse-weather conditions generated with OpenCV to cover distributions absent from real data. Evaluated across visibility levels to pinpoint exactly where the learned representation breaks down under distribution shift — findings that inform vision systems deployed in uncontrolled environments.
CNN vs. Vision Transformer Efficiency Study arXiv:2505.08259
An efficiency comparison of CNNs and Vision Transformers on Tiny ImageNet and DermaMNIST, examining the accuracy–compute trade-offs between architectures on constrained image-classification tasks. Co-authored with Amangeldi, Taigonyrov, and Mbonu.
Background
Education
MSc, Data Science
Nazarbayev University, Astana
BSc, Computer Systems Engineering
UET Peshawar, Pakistan
Experience
Research Assistant — Multimodal HAR
Nazarbayev University · signal + video fusion
Research Assistant — Climate ML
CMIP6 bias-correction with deep sequence models
Research Trainee & English Instructor
Computer vision research · independent teaching