Mohammed Ataiq
Alzahrani
Artificial Intelligence graduate turning research into production-grade systems — across machine learning, deep learning, NLP, speech, and computer vision. Selected among 15 of 8,000+ applicants for advanced AI training at KAUST Academy, with a focus on low-resource Arabic language technology.
Research-driven AI, engineered for the real world.
I work across the full AI lifecycle — from problem definition and data exploration through model development, system architecture, deployment, and monitoring. My work centres on Arabic and low-resource language technology: dialect corpora, speech systems, and adapting large models to serve languages they were never built for.
Where I've built
AI Engineer
- Leading the design and deployment of scalable, AI-driven solutions to solve complex business challenges.
- Owning the full AI lifecycle — from problem definition and data exploration through model development, system architecture, deployment, and monitoring.
- Building and optimizing machine learning and deep learning models with PyTorch and TensorFlow to production-grade reliability.
- Collaborating with cross-functional teams to translate business needs into high-impact AI systems.
- Driving technical decisions around performance, cost, and scalability trade-offs.
- Ensuring ethical AI practices and maintaining strong engineering standards.
AI Specialist
- Leading the design, optimization, and fine-tuning of deep learning models for high-performance STT and TTS systems.
- Architecting scalable AI systems for real-time speech applications.
- Managing large-scale audio and text data preprocessing and feature engineering.
AI Engineer
- Developed and optimized deep learning models for STT and TTS systems.
- Designed AI system architectures for speech applications.
- Conducted code reviews and ensured adherence to engineering best practices.
AI Scientist Intern
- Selected for the MISK Global Remote Traineeship Program, focused on Arabic NLP, STT, and TTS.
- Built and curated Arabic datasets for healthcare AI applications.
- Assisted in NLP model development and evaluation.
Artificial Intelligence Intern
- Completed two intensive Oxford programmes — Natural Language Processing and Deep Unsupervised Learning — placing in the top 0.18% of 8,000+ applicants.
- NLP: language models (N-grams, RNNs, LSTMs), transformer architectures, and LLMs (BERT, GPT), plus attention mechanisms, multimodal models, and prompt engineering, with hands-on PyTorch projects.
- Deep Unsupervised Learning: clustering, dimensionality reduction (PCA, t-SNE), generative models (GANs, VAEs, diffusion), contrastive learning, autoencoders, and StyleGANs.
- Enhanced Phi-3 Mini for Arabic and presented the work in a poster session at KAUST.
Projects
Hybrid ASR-T5 for French Speech Recognition
2025 – 2026A dual-encoder speech-recognition architecture that pairs a FastConformer acoustic encoder with T5's full encoder-decoder through a lightweight adapter — using the language model's encoder for linguistic enrichment before generation, unlike encoder-only (Whisper) or decoder-only designs.
Production ASR must stay accurate and robust across very different domains. Existing hybrids use only a language model's encoder (Whisper) or only its decoder — leaving linguistic capacity on the table.
Designed a dual-encoder hybrid: a FastConformer acoustic encoder feeds T5's complete encoder-decoder via a ~4M-parameter adapter, with LoRA adaptation and a dedicated silence-handling strategy to prevent hallucinations.
An 83.3% win rate (15/18 categories) over the production NeMo baseline trained on identical data, 3.9% validation WER (competitive with Canary and Voxtral), and an 18.6% relative WER reduction in the medical domain — with strong zero-shot transfer (22.78% WER with no medical training data). Being prepared for production deployment.
Arabic Sentiment Analysis with Full MLOps
Jan 2025An Arabic sentiment-analysis system built on fine-tuned BERT, handling data preprocessing, training, and monitoring — then deployed as a FastAPI app in Docker for real-time predictions.
Arabic sentiment is hard — rich morphology, many dialects, and little clean labeled data.
Fine-tuned BERT while handling data preprocessing, training, and monitoring, applying MLOps tools throughout: Prefect for orchestration, MLflow for experiment tracking, and Optuna for hyperparameter search.
Deployed a FastAPI app with Docker, serving sentiment predictions in real time.
Saudi Dialect Corpus & Translation Model
Jan – Dec 2024A comprehensive corpus of four Saudi dialects — Najdi, Hijazi, Gulf, and Southern — built by collecting and annotating linguistic data, paired with a model translating between dialects and Modern Standard Arabic.
Saudi dialects are underrepresented in NLP, with almost no parallel data to and from Modern Standard Arabic.
Collaborated on building a comprehensive corpus of four dialects by collecting and annotating linguistic data to capture their unique features, then trained a dialectal ↔ MSA translation model on it.
Achieved a BLEU score of 47, demonstrating strong performance for a low-resource setting.
Baseer — Historical Place Detection
Aug – Dec 2024A vision model on a modified CMT architecture that detects historical places, paired with an NLP RAG system (Sambalingo) and a chatbot interface that explains each detected site in detail.
Recognizing a historical site and explaining it well needs vision and language working together.
Developed Baseer on a modified CMT architecture for detection, then built an NLP RAG system (Sambalingo) with a chatbot interface to provide detailed explanations of detected sites.
Reached 97% accuracy — a higher score than the original CMT architecture while using just 0.01% of its parameters, showing that a far smaller model can beat the full one.
Augmenting Phi-3 for Arabic
Jul – Sep 2024Enhanced Microsoft Phi-3 Mini for Arabic by integrating LoRA, fine-tuning curated instruction data, and engineering a custom tokenization pipeline to handle Arabic script nuances.
Phi-3 Mini had weak Arabic support — the script and instruction-following were low-resource gaps.
Integrated LoRA, fine-tuned curated instruction data, and engineered a custom tokenization pipeline to handle the nuances of Arabic script.
Achieved robust Arabic language support, showcasing Phi-3's adaptability for low-resource language tasks.
Skills
Languages & Core
Domains
MLOps & Deployment
Methods
Tap any tool above to see how I use it.
Education & recognition
Education
University of Jeddah
Recognized for academic excellence (Excellence Certificates, 2020–2022). Coursework: deep learning, machine learning, NLP, artificial neural networks, recommender systems, multi-agent systems, optimization & regression.
Honors & Awards
Quran AI Competition, UJ — 2024
Proposed an NLP system for Quranic verse correction using Seq2Seq and CRF.
Tuwaiq AI Challenge — 2024
Built a model to predict program completion and optimize seat allocation.
Excellence Certificate — University of Jeddah
Awarded for an "Excellent" grade across two successive semesters (2021/2022).
Excellence Certificate — University of Jeddah
Awarded for an "Excellent" grade across two successive semesters (2020/2021).
Credentials
NVIDIA Certified Associate — Generative AI & LLMs
Generative AI, LLMs, prompt engineering, alignment, and deployment. Issued by NVIDIA.
AI Agents Fundamentals — Hugging Face
Foundations of building autonomous, tool-using LLM agents.
Introduction to Large Language Models — Google
Core concepts of LLMs, their uses, and prompt design.
KAUST Pre-Internship AI Training
72-hour on-site training — 40h foundations + 32h advanced AI techniques.
IBM AI Engineering Professional Certificate
Machine learning and deep learning with hands-on project experience.
Deep Reinforcement Learning Nanodegree
Q-learning, policy gradients, and actor-critic methods for robotics & autonomy.
Deep Learning Specialization — DeepLearning.AI
Five-course specialization covering deep learning concepts and applications.
Python 3 Programming — University of Michigan
Five-course specialization in Python 3 programming.
Data Engineering Foundations — IBM
Five-course specialization in data engineering principles and practices.
IBM Advanced Deep Learning Specialist
Deep learning architectures, optimization, and deployment with TensorFlow & PyTorch.
References
I highly recommend Mohammed Ataiq Alzahrani. I had the pleasure of supervising his senior project, during which he consistently demonstrated exceptional dedication, technical skill, and a strong work ethic. He impressed me with his ability to independently tackle complex problems, conduct in-depth research, and apply advanced AI techniques effectively. His project not only met academic expectations but stood out for its innovation and quality of implementation. He is a highly motivated individual who shows great promise in the field of artificial intelligence, and I am confident he will be a valuable asset to any organization he chooses to join.
Dr. Mohammad Alqahtani
Mohammed is one of my outstanding students. I was pleased to have him in my Deep Learning course. On several occasions I posed challenging questions and truly enjoyed the way he approached his answers. His critical thinking, solid reasoning, and strong background placed him among the top of his peers. His continued growth, along with his experiences at KAUST and Oxford, have shaped him into a well-rounded AI talent. I highly recommend him for any AI-related opportunity.
Dr. Hammam AlGhamdi
Full contact details available on request · Languages: Arabic (native) & English (professional working proficiency)