I am an engineer who builds things — medical imaging pipelines, protein-structure models, embedded control systems — and I go wherever the interesting technical problem is, not wherever the field boundary says I should stay.
I am passionate about solving interesting real life problems using computational methods.
First female-specific structural MRI brain template for the Kazakh population, built from 55 T1-weighted subjects. Led the pipeline design, probabilistic atlas construction, and cross-template validation. The early results were written up as a preprint and the final version is in the process of being submitted for a journal publication.
Preprint
Applied AlphaGenome biological deep learning framework to score candidate genomic variants for RNA and splicing effects, interpreting predictions across tissues as part of a computational biology literature review on ageing–cancer interplay.
GitHub
Designed and trained a custom 3-layer distance-weighted message-passing GNN over residue-level spatial graphs, built entirely from single-chain 3D structure — no partner chain, sequence alignments, or pretrained language models. Since interface residues are sparse, evaluated the model in a rank-based setting (PR-AUC, Precision@K, F1@K) with complex-level train/test splits to prevent leakage, benchmarking it against self-implemented logistic regression, random forest, and heuristic baselines on ~1,900 PDB complexes, with 3D case studies of top-ranked residues.
GitHub
Bare-metal C on ARM Cortex-M4 — interrupt-driven sensor fusion, complementary filter, PID control loop — debugged without any UART tooling.
GitHub
Fine-tuned ResNet-18/50 on 25,000+ DICOM chest X-rays with weighted loss for class imbalance; Grad-CAM heatmaps for interpretability. Best F1 0.6751, recall 0.7521, accuracy 0.84.
XGBoost and a stacking ensemble (SVM + XGBoost → RF meta) on imbalanced bank marketing data. Stacking ensemble reached F1 0.5606, AUC 0.9102, recall 0.7115.
Reproducible scikit-learn SVM pipeline on the UCI Breast Cancer dataset, tuned via GridSearchCV. 96% accuracy, 97% precision, 95% recall.