About Me
Data Scientist with a Master's in Data Science and hands-on experience delivering models in a banking environment. Built and evaluated classification models on imbalanced multilingual data, designed the evaluation framew…
Data Scientist with a Master's in Data Science and hands-on experience delivering models in a banking environment. Built and evaluated classification models on imbalanced multilingual data, designed the evaluation frameworks behind them, and worked end-to-end from raw tickets through to a deployed decision system. Comfortable across Python, SQL, statistical modeling and stakeholder-facing reporting.
Experience
Data Scientist Intern
<p>Raised classification Macro F1 to 0.82 by fine-tuning XLM-RoBERTa into a hierarchical pipeline — a 3-class router gating a 6-class minority model at a 0.80 confidence threshold — on 28k multilingual customer tickets (French, Arabic).</p><p> Corrected 84.7% single-class dominance through LLM-based augmentation and resampling, isolating a train-only export to close a data-leakage path that would have inflated test F1.</p><p> Designed the evaluation framework for a production RAG assistant, measuring faithfulness, answer relevancy, context recall and context precision over a 2,006-passage corpus and surfacing 3 systematic retrieval failure modes before deployment.</p><p> Analysed complaint volume and category distribution across 28k records to define the classification taxonomy and set per-class augmentation targets.</p><p> Built a banking assistant that autonomously resolved 70% of customer queries and routed the remainder to human agents through five confidence and keyword triggers.</p><p> Deployed models behind FastAPI and PostgreSQL, composing model output with deterministic domain rules, NLI validation and LLM arbitration for low-confidence cases.</p>
Computer Vision Intern
<p>Deployed a real-time facial recognition and driver fatigue detection system on Raspberry Pi 4, reaching 92% detection accuracy at 16 FPS.</p><p> Built the end-to-end pipeline from data collection through to on-device inference, tuning model size against embedded memory limits to keep recognition under 200ms per frame.</p><p> Integrated the vision system with a mobile application for automatic seat adjustment, closing the loop from detection to physical vehicle actuation.</p>