I am a
Computer Vision & AI engineer building multimodal medical AI, object detection, Object segmentation, object tracking, OCR's, VLMs, and research manuscripts ready for publication. Recent industry experience at CCRIPT Agency and Neuralogic(USA). Based in Islamabad.
Six months of computer vision engineering at CCRIPT Agency and Neuralogic (Jan–Jun 2026); research internship at NCAI (completed); degree work in parallel.
Manuscripts and systems ready for journal submission — multimodal medical AI, efficient video recognition, and clinical imaging. Lab work at NCAI complemented this with traffic CV and deployable pipelines.
ViT + Bio-ClinicalBERT with four-layer bidirectional cross-attention on 172,202 image–report pairs from two hospitals. 98.52% macro-AUC across 15 pathology labels; cross-hospital gap 1.30%; Grad-CAM interpretability.
Sparse local–global temporal attention with adaptive temporal pooling. 97.62% accuracy on UCF-101 at 71 GFLOPs — 44% fewer attention computations than full attention.
End-to-end system across 16 pathology categories: 95.20% test accuracy, 0.9921 weighted AUC-ROC. Grad-CAM explainability, severity prediction, and NIH ChestX-ray14 cross-validation.
ConvNeXt backbone with Kolmogorov-Arnold Network channel attention for CT stone detection. 99.13% accuracy, zero false positives on 346 test images; outperforms ResNet-50, DenseNet-121, VGG-16, and EfficientNet-B2. MixUp, two-stage fine-tuning, ablation study on GitHub.
More experiments and notebooks on GitHub. Preparing MS-oriented applications and open to academic collaborations in computer vision and trustworthy AI.
Stack I use daily for CV/ML delivery and research prototypes.
Client deliverables and research-grade builds below. Open-source work and notebooks are listed under “More repositories” and on GitHub.
End-to-end ML pipeline predicting urbanization and flood hazard from LULC satellite data. Six regression models per land-cover class (LOOCV); vectorized polynomial regression across 35M HEC-RAS pixels (Depth, Velocity, WSE). Near-perfect correlation (r = 0.99) between sprawl and worsening flood risk. Delivered GeoTIFF maps, ArcGIS overlays, notebooks, and technical documentation.
End-to-end SER system: merged RAVDESS, CREMA-D, TESS, and SAVEE (62,000+ samples), 234 acoustic features per clip (MFCC, Mel, Chroma, Spectral Contrast, Tonnetz). Compared SVM, 1D-CNN, CNN-LSTM, and CNN-LSTM with channel + spatial attention — novel attention variant targeting 93–97% across eight emotions.
Real-time vehicle detection with signal timing from lane density using YOLOv11.
YOLOv8 plus PaddleOCR for plate text and CSV export.
YOLOv11 on MRI-style scans for tumor and dead-cell regions.
Classify whether appliances are on or off from camera frames with YOLO.
Multi-class YOLOv11 detectors for emergency and rickshaw traffic.
Each line is a one-line summary. To read code, notebooks, or train configs, open the repo on GitHub — everything lives under github.com/daud-shah.
Want the full picture? Visit all repositories on my GitHub profile.
Have a project in mind or want to collaborate on AI research? Let's connect and build something incredible.
Islamabad, Pakistan