Ziad El Sayed
Portrait of Ziad El Sayed.

Ziad El Sayed

Email: ziadelsayed at nyu.edu

About Me

I am an AI researcher and engineer based in London, with an MSc in Machine Learning from University College London, awarded with distinction, and a BSc in Computer Science from NYU Abu Dhabi. My work has been published in ACM Computing Surveys and ACL SRW.

I am currently most interested in studying and mitigating demographic bias in machine learning models, and in the growing presence of AI in mental health and well-being.

Research Interests

Broadly, I am interested in the social consequences of AI systems. I also work on technical questions in interpretability, memory, and representation. Please feel free to reach out if any of this resonates.

  • AI and Social Inequality: I am interested in measuring how language and vision models respond to a person's identity, whether stated or inferred from how they communicate or appear, and how this shapes the information and decisions they receive.
  • Help-Seeking and Conversational AI: People are increasingly turning to LLMs for companionship, emotional support, and medical advice. I am interested in whose idea of care these models are built around, whether they tell people what they want to hear (i.e. sycophancy), and what that reliance displaces.
  • AI Safety and Interpretability: I am interested in mechanistic interpretability, and have worked on the faithfulness of sparse autoencoders. More broadly, I am interested in the role interpretability can play in detecting and mitigating misalignment.
  • Memory in AI Agents: I am interested in the retrieval architectures that keep agents coherent and factually consistent over long interactions, especially as they are deployed in more sensitive settings.
  • Graph Neural Networks: I am interested in solutions that leverage graph neural networks (GNNs) to model data with inherent graph-like structures. My research investigates the use of GNNs in advancing integrated circuit design, reliability, and security.

News

  • Gave an invited talk at Koç University’s KUIS AI Center on OMem, my work on long-term memory for LLM agents
  • Started at Sadel Group in London as an AI Engineer
  • My first-author survey on graph neural networks for hardware design, security, and reliability is out in ACM Computing Surveys!
  • Represented Oxtractor at Google Cloud’s CTO Connect with DeepMind in London
  • Finished my MSc in Machine Learning at UCL, graduating with distinction
  • Our paper FaithfulSAE was accepted to the ACL 2025 Student Research Workshop
  • Joined Oxtractor, an Oxford University Innovation venture, to work on memory retrieval for real-time agentic systems
  • Started working on mechanistic interpretability at UCL, on the faithfulness of sparse autoencoders
  • Graduated from NYU Abu Dhabi with a BSc in Computer Science and minors in Applied Mathematics and Business Studies
  • Presented my research on multimodal and demographic-based bias mitigation at the NYU Abu Dhabi Capstone Festival
  • Awarded over $3,400 to spend the summer researching ML for hardware security at NYU
  • Returned to the NYUAD Hackathon for Social Good as an organiser
  • Served as a teaching assistant for Language of Computers (CADT-UH 1013), an introductory programming course at NYU Abu Dhabi
  • My team won Best Application of Social Good at the NYUAD Hackathon for Social Good with qSa’id, a hybrid quantum-classical ML platform that hit 95.1% accuracy screening for autism

Selected Publications

  1. Flow diagram: an input circuit netlist is converted to a graph, passed through GNN layers to produce an embedding, then through linear layers to downstream EDA, hardware reliability and hardware security tasks.

    Graph Neural Networks for Integrated Circuit Design, Reliability, and Security: Survey and Tool

    Ziad El Sayed, Zeng Wang, et al.

    ACM Computing Surveys, 58(4), 1-44, 2025

    Paper: Graph Neural Networks for Integrated Circuit Design, Reliability, and SecurityGitHub: GNN4IC repository

  2. Across seven language models, sparse autoencoders trained on the model’s own generated data (Faithful) have a lower fake feature ratio than those trained on Pile or Fineweb for GPT2-small, Gemma 2B and LLaMA 1B, 3B and 8B, but not for Pythia 1.4B or 2.8B.

    FaithfulSAE: Towards Capturing Faithful Features with Sparse Autoencoders without External Datasets Dependency

    Seonglae Cho, et al., including Ziad El Sayed

    Proceedings of the ACL Student Research Workshop, 297-314, 2025

    Paper: FaithfulSAE