Zeyad Ezzat

- Search Platform Architecture: Led the redesign of a production AI search platform, simplifying multi-index orchestration and enabling rapid integration of retrieval capabilities across multiple products.
- Retrieval Optimization: Designed advanced document chunking, segmentation, and preprocessing strategies that increased Recall@10 by 10 percentage points while maximizing downstream LLM context

- Language Models: Continued pretraining transformer language models on billions of Arabic tokens to improve semantic representations for downstream NLP applications.
- Distributed Training: Managed distributed training workflows for transformer models ranging from millions to 70B parameters.
- Efficient Deployment: Reduced inference cost using PEFT, LoRA, BitsAndBytes, and AWQ while maintaining m

- Distributed Training: Built scalable transformer training pipelines across multi-GPU and multi-node environments using DeepSpeed, Slurm, and Kubernetes.
- Training Optimization: Improved GPU utilization through DeepSpeed ZeRO optimization, optimizer offloading, and mixed precision training.
- Dataset Engineering: Developed preprocessing pipelines for structured and unstructured datasets supporti

Teaching Assistant in the School of Information Technology and Computer Science.
- Delivered tutorials and lab sessions, both online and in person, for undergraduate courses including Machine Learning, Computer Vision, Image Processing, Natural Language Processing, and Data Mining.
- Designed lab exercises and programming assignments, and mentored students through hands-on implementation tasks.
-

- Tokenizer Development: Designed custom Byte-Pair Encoding (BPE) tokenizers and pretrained masked language models from scratch for Arabic NLP.
- Semantic Representations: Evaluated classical and neural text representations, including TF-IDF, word embeddings, and transformer encoders for downstream NLP tasks.

Arabic NLP research on dialect identification for the NADI 2020 shared task (Nuanced Arabic Dialect Identification).
- Ranked 5th in the NADI 2020 shared task.
- Built and compared classical machine learning baselines and deep learning models, including fine-tuned BERT.
- Ran and analyzed experiments to systematically improve the team's leaderboard score.
- Published "Arabic Dialect Identification

Course work:
Image Processing and Computer Graphics:A+
Cloud Computing: A
System Engineering: A
Machine Learning and Data Analysis: A-
Advanced Big Data: A-
Computer Vision: A-
Deep Learning: A-
Intro to Big Data: B
Project Management : B

Sep 2020

Oct 2020

Jul 2020

Jun 2020