At JLL, I built a PySpark entity resolution pipeline for JLL Azara that standardized 25,000+ multilingual company-name records. It reduced the duplicate entity rate from ~18% to <2% and saved ~3–4 person-weeks per refresh cycle.
At JLL, I also developed modules for a commercial lease-renewal scoring pipeline, using XGBoost to flag at-risk leases up to 18 months before expiry. The classifier achieved an AUC of 0.80, and its predictions fed a downstream NPV optimization engine.
In my independent AI research, I fine-tuned AraBERT for Arabic dialect identification, achieving 86% macro-F1 across 26 city-level labels. I also built FinTrace, a financial question-answering system that achieved 96.66% Hit@1 for evidence ranking, and a multi-agent RAG system for SEC financial disclosures.

