Abdelrahman Abouzeid
Software engineer by day, independent AI researcher by night. B.S. in Computer Science from Rice University.
Work
NestMed Health Tech Startup
June 2024 – PresentSenior Software Engineer
Building and scaling NestMed's HIPAA-compliant API platform, serving 100+ clinicians and home health agencies. Working across the stack with Python, Django, AWS ECS, and Temporal.io — from autoscaling infrastructure to automated insurance validation workflows.
Telda Sequoia-backed fintech
May 2023 – June 2024Software Engineer
Lead engineer for onboarding and customer experience in the Investments division. Built the backend for in-app stock trading on the Egyptian market, an in-app chat feature adopted by 33,000+ users, and led the "Spaces" savings product end-to-end.
Meta (Facebook)
Sep 2022 – Jan 2023Software Engineer
Shipped features across WhatsApp, Instagram, and Messenger, collaborating with teams globally on backend and frontend work.
Arista Networks
May 2021 – Aug 2021Software Engineer Intern
Enhanced the EOS CLI and fixed a major bug in storage device information collection, tested across 50+ hardware devices.
Research Interests
Outside of work, I independently research foundational models and the infrastructure behind them.
Foundational Models
Architecture design, training dynamics, and scaling properties of large language models.
LLM Infrastructure
Efficient training pipelines, distributed systems for model parallelism, and inference optimization.
Training Optimization
Novel optimizers, learning rate scheduling, and techniques for stable and efficient model training.
Publications
Does Your Optimizer Care How You Normalize? Normalization-Optimizer Coupling in LLM Training
April 2026Preprint · arXiv
Investigates whether normalization techniques and optimizers should be treated as independent design choices in LLM training. Through factorial experiments at 1B parameters, demonstrates that Dynamic Erf normalization performs significantly worse under Muon's spectral-norm dynamics compared to AdamW, and proposes solutions including EMA-blending that recover much of the lost performance.
Contact
Feel free to reach out if you'd like to discuss research or collaboration opportunities.