AI Researcher & Software Engineer
I work on Arabic NLP, multimodal evaluation, and large language models - from dialectal generation and OCR to hallucination detection and human-preference benchmarking.
Arena-style human evaluation platform for LLMs across 11 Arabic dialects - 1,100+ pairwise comparisons over 67 models, ranked via Bradley–Terry (ESS = 582).
CER = 0.0819, WER = 0.2588 on blind per-line evaluation - ranked 1st in that setting.
DER = 10.56%, WER = 34.47%, SER = 79.88% - 5th of 7 teams in the KSAA-2026 shared task.
Presented at 1st Research Conference for Students of Makkah Region Universities (RCUSM), Jeddah.
Ancestry-stratified calibration and explainability analysis for machine learning-based warfarin dosing models.
Zero-shot LLM classification with self-consistency sampling reached 0.7013 hierarchical macro F1, ranking 3rd of 18 teams.
Dialect-aware sentiment classification across Moroccan, Egyptian, Jordanian, and Saudi Arabic - MARBERTv2 achieved the strongest result at 94.07% macro F1.
CAMeLBERT cross-encoder hallucination detector achieved 0.7596 ROC-AUC, ranking 3rd of 7 teams; temperature scaling and MC-Dropout deferral further improved calibration and reliability.
Retrieval-augmented in-context learning ensemble for Arabic stance detection achieved Favg2 of 0.8719 (Track 1, ranked 3rd) and 0.9016 (Track 2, ranked 7th).
Cross-backbone ensembling for Arabic dialectal sentiment-polarity swap reached 96.78 development accuracy; best official test score was 0.7597.
Cause-diagnostic Arabic visual hallucination benchmark for multimodal LLMs, evaluated across four models.
4-model Arabic encoder ensemble (AraBERT, AraBERT-Large, ARBERTv2, MARBERTv2) for detecting cultural hallucinations - 0.9647 macro F1 on 457 validation examples, with Arabic explanations generated for all 163 detected hallucinations.
Reproducible, fairness-aware pipeline for warfarin dose prediction - leakage-free preprocessing, ablation across 9 model architectures, leave-one-ancestry-out evaluation, conformal prediction calibration, SHAP explainability, and a FHIR R4 clinical decision-support interface.
King Abdulaziz City for Science and Technology (KACST) · Riyadh, Saudi Arabia (On-site)
NAMAA Community · Riyadh, Saudi Arabia (Remote)
King Khalid University · Abha, Saudi Arabia (Remote)
Alfaisal's AI Center · Riyadh, Saudi Arabia (On-site)
Umm Al-Qura University · Jeddah, Saudi Arabia (Hybrid)
MawHub · Riyadh, Saudi Arabia (Remote)
Synapse · Jeddah, Saudi Arabia
Lucidya · Riyadh, Saudi Arabia (Remote)
Smart Methods · Jeddah, Saudi Arabia (On-site)
Specialized Med AI Lab (SMAIL) · Riyadh, Saudi Arabia (Remote)
Global Students Research Conference (GSR) · Riyadh, Saudi Arabia (Remote)
Global Students Research Conference (GSR)
Umm Al-Qura University
AL-YAMAMA Company
AL-YAMAMA Company
Ministry of Sport
Ashour Hotel
Ashour Hotel
Live Product · Built at Alfaisal's AI Center
Live Product · Built at Umm Al-Qura University
Live Product
Live Product · Built at Lucidya
Live Product · Built at Synapse
Live Product · Specialized Med AI Lab (SMAIL)
Paper (arXiv) · Aug 2025 · First-author, RCUSM Jeddah
NAMAA Community · 2026
Model on Hugging Face · Apache 2.0
MIT License
MIT License · Jun–Aug 2025
MIT License
Kaggle Notebook
Kaggle Notebook
Kaggle Notebook
Kaggle Notebook
Kaggle Notebook
Kaggle Notebook
Kaggle Notebook
Re-implementations of influential AI/ML papers from scratch - focused on understanding experimental design, methodological rigor, and the gap between paper and practice. Work in progress.
Umm Al-Qura University · Makkah, Saudi Arabia
GPA: 3.87 / 4.00
Associated with Umm Al-Qura University
Associated with Umm Al-Qura University
Associated with الأساليب الذكية Smart Methods