Ahmed Hassan

Ahmed Hassan

AI Engineer and Researcher specializing in Computer Vision and NLP.

Publications

Here you can find a list of my academic publications (published and submitted). I am passionate about sharing my work and collaborating with the research community.

DAgent: A Multi-Agent System for Device-Aware Assistance

In this work, we proposed DAgent, a modular multi-agent system designed for personalized, device-aware assistance. The system architecture includes specialized agents for tracing, retrieval-augmented generation (RAG), and coding, all orchestrated by a multi-agent OS assistant. Our evaluations showed strong performance in correctness, completeness, and clarity. Ablation studies highlighted the critical role of the Coding Agent, demonstrating the power of multi-agent collaboration for complex, real-world tasks.

Computational Analysis of Media Bias in The Guardian’s Coverage of the Israeli-Palestinian Conflict

This paper presents a multi-method NLP analysis of media framing in The Guardian’s coverage of the Israeli-Palestinian conflict over several decades. Using transformer-based models and word embeddings, we uncovered systematic biases in language use. For instance, our findings revealed that fear-related language was 13.5% more prevalent in contexts related to Israel, whereas trust-related language was 9.5% more prevalent in contexts related to Palestine, providing a quantitative look at sentiment patterns in journalism.

ALEX-GYM-1: A Hybrid 3D Pose & Vision Model for Automated Exercise Evaluation

Current automated exercise evaluation systems often rely on a single modality, like pose estimation. In ALEX-GYM-1, we developed a novel multi-modal architecture that integrates both 3D CNN vision-based and pose-based pathways. This hybrid approach allows the model to capture a richer set of features, leading to a 30% reduction in Hamming Loss compared to single-modality methods and a 79.5% improvement over pose-only models, setting a new standard for exercise form evaluation.

TypePlus: A Deep Learning Architecture for Keystroke Authentication

We designed TypePlus, a lightweight, non-transformer architecture for free-text keystroke authentication. The model leverages weighted attention pooling and keycode embeddings to effectively capture individual typing patterns. Our approach achieves a state-of-the-art 2.86% Equal Error Rate on the Aalto University Keystroke Dataset, demonstrating that highly accurate biometric security can be achieved with efficient, non-transformer models.

PPE-Det: Evaluating Lightweight Object Detection Models for Edge-Based Safety Monitoring

Real-time safety monitoring in industrial environments requires efficient models that can run on edge devices. For this project, we first created a novel dataset of 5,000 images of personal protective equipment (PPE). We then benchmarked several lightweight (sub-3M parameter) object detection models on a Raspberry Pi 5. Our analysis found that YOLOv9t and YOLOv11n achieved the optimal trade-off between accuracy and efficiency, providing a clear recommendation for real-time edge deployment in industrial safety systems.

DeepCat: A Deep Learning Approach to Understand Your Cat’s Body Language

This project introduced a mobile-friendly deep learning system designed to interpret a cat’s emotional state by analyzing its body language from images. Trained and evaluated on over 10,000 images, the system uses key visual markers to recognize emotions. It achieved 97% accuracy for eye detection, 85% for tail positions, and 84% for mouth configurations, demonstrating a practical and accessible application of computer vision for animal welfare and human-pet interaction.