Hi, I'm Raoof.

The pipeline classifies input radiographs into the correct anatomical category and separately predicts normal vs. abnormal, using a dedicated preprocessing stage with specialized feature-extraction modules feeding into convolutional and Squeeze-and-Excitation blocks. Iterated through roughly 16 model variants to get there — working through focal loss, label smoothing, and both positive and negative augmentation strategies, including region-specific tuning, to handle class imbalance and subtle abnormality signals. Final model reached 97.37% accuracy on anatomical classification and 89% on abnormality detection, with ROC-AUC 0.94, sensitivity 0.86, specificity 0.89, and Cohen's kappa 0.77. Delivered as a complete model and architecture handoff to a healthcare research client.

MSc dissertation using transfer learning on MobileNet against the Adience dataset (19,370 images) to investigate lightweight age estimation for edge deployment. Exploratory analysis surfaced key dataset issues — age-group imbalance, gender variance, and labeling anomalies — that shaped the preprocessing and modeling approach. Iteratively diagnosed and corrected underfitting and overfitting across model variants using Augmentor-based data augmentation, Stratified K-fold cross-validation, and L2 regularization, converging on a final architecture that reached 88.88% accuracy on binary (minor/adult) classification and 74.33% on six-group age classification — evaluated against a full metric suite (ROC-AUC, not accuracy alone) to give a rounded view of performance rather than a single headline number.

Built during my role as Senior Trainer (Cognitive Science) at the Marík Institute of Computing, Artificial Intelligence, Robotics, and Cybernetics. Designed a Python/Streamlit cueing interface presenting randomized motor-imagery trials (left, right, forward, backward, rest) to build a balanced dataset, recorded synchronized EEG/EOG/EMG via a Kaptics acquisition kit at ~135–142 Hz, and cleaned the signal with ASR (artifact subspace reconstruction) and ICA (independent component analysis). Also built real-time demos integrating BCI output with live cognitive experiments.

A modular, asynchronous multi-agent system for automated research and executive summarization using Llama 3.1 via the Groq API — built with asyncio orchestration and Pydantic-validated data passed between agents so each hand-off stays typed and predictable. Two researchers gather sources in parallel while a manager agent reconciles overlaps and synthesizes everything into a single polished Markdown report.

Delivered as an independent freelance engagement: a PyTorch-based object detection pipeline trained to flag ticket status in real time, reaching 95%+ accuracy and translating directly into fewer evaded fares — a clear revenue-protection outcome, not just a model demo.

For a travel/hospitality client, K-Means clustering split the customer base into distinct behavioral cohorts, while an XGBoost model flagged churn risk within each segment — letting marketing target the right offer to the right cohort instead of running one blanket campaign, and driving a 15% increase in repeat bookings.