Ziad El Sayed

Projects

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

2025

Graph neural networks (GNNs) have excelled in learning tasks for graph-structured data, including social networks, biology, and circuits. Their success in circuit-related tasks has fueled growing interest in applying GNNs to IC design. This project presents the first comprehensive survey of GNN applications in IC design, covering electronic design automation (EDA), reliability, and security. It introduces a generic GNN application flow spanning graph conversion, feature extraction, architecture selection, and task formulation, along with GNN4CIRCUITS, a customizable Python tool for researchers. The paper also explores open challenges and future research directions. Published in ACM Computing Surveys.

Two chat transcripts side by side. Without persistent memory the assistant forgets a stated celiac diagnosis and later suggests a tortilla. With persistent memory it recalls the constraint and suggests a gluten-free tortilla.

OMem: Dynamic Hybrid Retrieval for Long-Term Memory in LLM-Based Agentic Systems

2025

For my MSc thesis, I developed OMem, a hybrid memory framework for agentic LLMs that integrates dense semantic retrieval, sparse BM25 scoring, reciprocal rank fusion, and cross-encoder reranking into a unified, modular pipeline. Benchmarked on LongMemEval and LoCoMo, it scored 76.6% and 75.52%, above the reported results of leading memory systems, improving coherence, factual consistency, and adaptability in persistent LLM agents. Conducted with Oxtractor.

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

2025

Sparse Autoencoders (SAEs) have emerged as a promising solution for decomposing large language model representations into interpretable features. However, Paulo and Belrose (2025) have highlighted instability across different initialization seeds, and Heap et al. (2025) have pointed out that SAEs may not capture model-internal features. These problems likely stem from training SAEs on external datasets—either collected from the Web or generated by another model—which may contain out-of-distribution (OOD) data beyond the model’s generalisation capabilities. This can result in hallucinated SAE features, which we term “Fake Features”, that misrepresent the model’s internal activations. To address these issues, we propose FaithfulSAE, a method that trains SAEs on the model’s own synthetic dataset. Using FaithfulSAEs, we demonstrate that training SAEs on less-OOD instruction datasets results in SAEs being more stable across seeds. Notably, FaithfulSAEs outperform SAEs trained on web-based datasets in the SAE probing task and exhibit a lower Fake Feature Ratio in 5 out of 7 models. Overall, our approach eliminates the dependency on external datasets, advancing interpretability by better capturing model-internal features while highlighting the often neglected importance of SAE training datasets.

Two rows of photographs from the image sets built to extend MMBias: the African target group above, the South Asian target group below.

Bias Mitigation using Functional Inequalities for Regularization

2024

For my undergraduate capstone, I assessed demographic bias in vision-language models. To achieve wider demographic representation in the MMBias dataset, I collected images for two additional target groups, African and South Asian, through the Flickr API, following the original creators’ pipeline: retrieving the 1,000 most relevant images from 10 search keywords, filtering out irrelevant or inappropriate images with three human annotators to minimise the effect of individual biases, and randomly sampling 250. I then built an open-source, reproducible implementation of two bias assessments, cross-modal zero-shot classification and effect-size bias scores, and found that OpenAI’s CLIP exhibits bias against the added groups. Separately, I developed a method for mitigating multimodal bias using regularization based on functional inequalities. On Colored MNIST, it reached 97.64% mean accuracy (± 0.71) on grayscale digits, 1.4 percentage points above the previous best of 96.24%.

Three screens from the Spotify redesign prototype: a library of playlists, a playlist with its track list and playback bar, and a podcast page showing episodes by season.

Spotify Reimagined

2022

Redesigned Spotify to be more user-centered and accessible, implementing various human computer interaction principles. Conducted user experience research and used it to create a UI wireframe and prototype using Figma.

The qSa’id team in front of the NYUAD International Hackathon for Social Good sign.

qSa’id: Quantum-ML-Assisted Diagnostic and Treatment Access Platform for Autism

2022

The main aim of qSa’id is to cater to the needs of people with autism by providing a quantum approach for early screening and allocating resources efficiently among those who require them. qSa’id involves an efficient hybrid quantum-classical ML approach towards improving automated screening, and a parallel solution using QUBO to optimize placements of specialized treatment centers in a health system. Winner of Best Application of Social Good as part of NYUAD’s 2022 Hackathon for Social Good.