Capabilities
This page avoids self-scored percentages. Each capability is connected to work that demonstrates how it was applied.
| Capability | Applied tools and methods | Evidence |
|---|---|---|
| Reliable RAG | Hybrid retrieval, BM25, TF-IDF, text normalization, citations, controlled generation | GovRAG Copilot, Asthma RAG |
| Model evaluation | Baselines, repeated seeds, ROC/PR, macro F1, class-level errors, cross-dataset testing | Deepfake robustness, Malicious URLs |
| Computer vision | PyTorch, CNN transfer learning, YOLOv8, OpenCV, privacy blurring, DCT features | PPE detection, Cashew disease |
| NLP and text analytics | Arabic/English preprocessing, TF-IDF, Word2Vec/AraVec, LSTM/BiLSTM, transformers | Arabic sentiment, Galaxy sentiment |
| Data and decision modeling | Pandas, feature engineering, imbalance review, segmentation, backtesting, interpretation | Retail analytics, Market regimes |
| Prototyping and implementation | Python, Git/GitHub, testing, Gradio, Streamlit, reusable project structures | Selected work |
Working principles
- Define the decision and user before choosing the model.
- Compare against clear baselines and report tradeoffs, not one metric.
- Treat citations, privacy, monitoring, and failure modes as system requirements.
- Keep repositories inspectable, reproducible, and documented.