I co-conceived and co-led AraHarMeme, a multimodal Arabic benchmark for harmful meme detection, and presented the work at JCSSE. I designed its annotation protocol and evaluated performance across dialectal strata to examine robustness, generalization, and bias in generative models.
At Institut Jean Lamour, University of Lorraine, I developed AI models to analyze synchrotron XRD data and predict steel phase fractions from diffraction signals. I generated synthetic training data with under 1% reconstruction error and used SHAP and LIME to interpret model decisions against physical domain knowledge.
At RMA ARZ Assurance, I designed and deployed an Agentic HybridRAG system for semantic querying of structured and unstructured insurance data. I also developed document-processing pipelines and multi-agent workflows for recommendations and multilingual processing, delivering an AI tool adopted by internal teams.
My projects include a real-time crypto anomaly detection system using Kafka, PySpark, and Isolation Forest, and a churn prediction pipeline that improved high-risk customer detection from 0.54 to 0.72 recall. I also built analyses and applications for causal inference, financial risk, marketing attribution, and emissions forecasting.
