Civil Engineer & AI Enthusiast
I build أبني
From SaaS platforms to intelligent bots — here's what I've built.
A semantic search engine for Arabic poetry — search 6.57M classical and Nabati verses by meaning, not keywords. Built on a 384-dim Arabic embedding model and an HNSW index, and benchmarked on Fann-or-Flop (EMNLP 2025) at ~22× random.
An AI agent that searches every Saudi government source and compiles the full licensing requirements for any business — documents, fees, steps, and official links — in one bilingual report.
An AI system that identifies the reciter, surah, and ayah from any Quran recitation clip in about 20 seconds. Recognizes 50 reciters (99% validation accuracy) and matches against all 6,236 ayahs. Available as a Telegram bot and a web demo.
An AI-powered logistics management platform for Saudi businesses. Features smart shipment search, real-time risk analysis, ZATCA-compliant e-invoicing, and a full auditor portal.
A full-stack Arabic-first property management system for Saudi landlords. Manage units, tenants, contracts, maintenance, utility bills, and generate financial reports — all in one place.
A compliance platform for tracking and reporting Saudi local content (LC) percentages across procurement and projects — aligned with Vision 2030 requirements.
A desktop-based intelligent logistics management system with a full GUI. Handles shipments, inventory, suppliers, and generates reports — built for operations teams without internet dependency.
An AI Telegram bot that diagnoses plant diseases from a single photo. Powered by 18 specialized AI models trained on Saudi agriculture crops — available 24/7 and fully in Arabic.
A fine-tuned AI model that translates English text into four authentic Saudi Arabic dialects: Najdi, Hijazi, Eastern, and Southern. Trained on 31,000+ sentence pairs with a BLEU score of 30.04.
Published and ongoing research in Arabic NLP and language models.
A comprehensive adversarial robustness benchmark of Arabic BERT models against five attack types. Reveals a critical accuracy–robustness trade-off: MARBERT resists diacritical attacks (>92% retention) but sacrifices baseline accuracy. The mechanism is traced to tokenizer construction.
Read paper →Challenges the assumption that newer, larger LLMs are universally superior. Demonstrates that modern generative LLMs (Qwen 2.5, Llama 3.1) suffer from severe semantic anisotropy in Arabic — while the 110M-parameter AraBERT outperforms 8B+ LLMs for semantic retrieval tasks.
Read paper →A benchmark study comparing generalist LLMs (Llama-3, Qwen-2.5) against specialized encoder models (CAMeLBERT) on noisy Arabic sentiment analysis. Qwen-2.5 achieves 91.5% accuracy — but with 28% higher latency than Llama-3.
Read paper →A published study on activating Building Information Modeling (BIM) in construction project management using artificial intelligence tools.
Read paper →A published study on how digital and strategic transformation drives governance excellence in non-profit organizations.
Read paper →A published study on strengthening AI governance through ethical handling of sensitive data, with applications in text classification and differential privacy.
Read paper →Have a project in mind, or need access to project files? Feel free to reach out.