Asalah
Live Product · Built at Alfaisal's AI Center
Generative AI evaluation platform benchmarking cultural fidelity of image-generation models across Saudi cultural domains — 10 models, 1,420 generated images, 142 prompts.
Live Product · Built at Alfaisal's AI Center
Generative AI evaluation platform benchmarking cultural fidelity of image-generation models across Saudi cultural domains — 10 models, 1,420 generated images, 142 prompts.
Live Product · Built at Umm Al-Qura University
Arena-style human-preference evaluation platform for Arabic LLMs — 1,100+ pairwise comparisons across 67 models and 11 dialects, ranked via Bradley–Terry.
Live Product
Website for the 1st Research Conference for Students of Makkah Region Universities (RCUSM).
Hugging Face · May 2026
System for detecting cultural hallucinations in Arabic LLM outputs across dialectal, religious, and regional contexts. Achieved 0.9647 macro F1 using a 4-model Arabic encoder ensemble.
Lucidya · Nov 2025 – Present
Engineering quality and delivery analytics system. Automation layer integrated into daily engineering workflows to enforce PR quality and documentation completeness.
Paper (arXiv) · Aug 2025 · First-author, RCUSM Jeddah
LoRA fine-tuning of ALLaM-7B for Saudi dialect generation (Hijazi & Najdi). Dialect-token variant: 84.21% Saudi alignment, 6.21% MSA leakage (base 47.97% / 32.63%).
Parameter-efficient Arabic HTR for historical manuscripts. QLoRA fine-tuning with DoRA and RSLoRA; Linear+Boost weighted ensemble. 1st on per-line evaluation (CER 0.0819, WER 0.2588), 3rd on corpus-wide.
NAMAA Community · 2026
Systematic evaluation of 18 diacritization models (Seq2Seq, token classification, decoder LLMs, ASR+Text) for Arabic speech dictation. 5th place on the official leaderboard (DER 10.56%, WER 34.47%, SER 79.88%).
Model on Hugging Face · Apache 2.0
Experimental Arabic OCR for NakbaNLP 2026: LoRA fine-tuning of Qari-OCR (Qwen2-VL-2B) with DoRA and RSLoRA. CER 0.2635 on blind test.
MIT License
Multi-Layer Perceptron in plain Python with no deep learning frameworks — gradient descent, backpropagation, and training loop by hand. 98.33% training accuracy and 100% test accuracy on Iris.
Flask web app that generates ATS-friendly PDF resumes via LaTeX compilation in Docker. Drag-and-drop section ordering, dynamic entries for experience, education, projects, and publications.
MIT License · Jun–Aug 2025
Robotics engineering internship deliverables: 5-DOF robotic arm design, Arduino/ESP32 integration, computer vision with OpenCV, voice assistant with STT/TTS/LLMs, ROS2 Turtlesim.
MIT License
Python chat application with real-time messaging via XML-RPC. GUI with customtkinter, multithreaded server for concurrent clients, multiple rooms, and online user listing.
Kaggle Notebook
Arabic Alphabets Sign Language classification over 97% accuracy. Computer vision pipeline for recognizing Arabic sign language gestures from the ArASL dataset.
Kaggle Notebook
Fruit image classification using EfficientNet with accuracy above 99.91%. Transfer learning and fine-tuning for robust multi-class recognition.
Kaggle Notebook
K-Means clustering from scratch — initialization, iteration, and evaluation with visualizations on sample datasets.
Kaggle Notebook
Conditional Deep Convolutional GAN for generating face images by gender using the CelebA dataset.
Kaggle Notebook
DDPM applied to the Stanford Cars dataset. Forward and reverse diffusion process for high-quality car image generation.
Kaggle Notebook
NLP sentiment classification end-to-end: preprocessing, embedding, and classifier for sentiment analysis.
Kaggle Notebook
Vanilla GAN trained on MNIST — generator and discriminator from scratch for digit image generation.