Research

This work presents a hardware-aware framework that integrates Quantization-Aware Training (QAT) with Layer-Wise Mixed-Precision Quantization (MPQ) to enable sub-millisecond, microjoule inference for indoor environment classification on the MAX78002 microcontroller.

In this work, we manually verified and labeled a substantial chunk of the public portion of the Muharaf dataset, enhancing its applicability for supervised learning and writer identification. We developed an end-to-end CNN-based DL system with attention mechanisms for line-level writer identification in historical Arabic handwritten manuscripts using this labeled Muharaf Dataset, accommodating up to two authors per line.

This work presents a deployable, lightweight machine learning (ML) framework for Intrusion Detection Systems (IDS) designed to operate across edge, fog, and cloud layers of the IoT stack

We comprehensively review the quantization landscape, specifically for microcontroller-class (MCUs) devices. The survey covers advanced quantization techniques, hardware platforms as part of three families (ARM-based, RISC-V-based and NPU-based), software toolchains, and applications.

We study the effect of quantization-aware-training (QAT) on two SOTA spectrum sensing models - DeepSense and ParallelCNN. Models are deployed on a Sony Spresense for hardware evaluation.

This work explores the deployment of deep learning models on resource-constrained edge devices to monitor student engagement in real time, with an emphasis on efficiency and privacy.

The survey offers a systematic, data-driven review of diffusion-based generative models for synthetic sequential data published between 2023 and 2025, organizing existing approaches by sequence dimensionality, cardinality, and structural complexity, and examining recent work in terms of architectures, training strategies, and evaluation practices. It identifies common challenges, highlights gaps in current research, and outlines future directions to improve applicability, robustness, and cross-domain generalization.

This survey provides a systematic, up-to-date review of CRL from 2022–2025, outlining the field’s conceptual foundations, proposing a taxonomy of recent methods by core mechanisms, and critically analyzing current benchmarks, metrics, and evaluation protocols, with particular focus on open challenges such as unified evaluation criteria, ecological validity, and assessment in multi-agent and interactive settings.

This project implements a Continual Reinforcement Learning (CRL) framework for mobile robot navigation using the e-puck robot in Webots. We design a unified agent trained sequentially across multiple tasks—maze navigation, line following, and obstacle avoidance—using Soft Actor-Critic (SAC). The setup evaluates the agent’s ability to learn new behaviors while retaining previously acquired skills.

This was my MLR503: Data Mining and Knowledge Discovery Course Research Project. We developed an end-to-end deep learning-based handwritten text recognition (HTR) system for Arabic script leveraging the KHATT Dataset. To further enhance recognition accuracy, we incorporated KenLM for post-processing.

This was my B.Sc. in Computer Engineering Senior Design Project, focused on spectrum sensing and allocation using a low-complexity deep learning-based (CNN) spectrum sensing algorithm. The project involved developing and quantizing the CNN model, which was deployed on hardware for real-time operation. The solution was demonstrated both in simulation and on hardware, utilizing a Raspberry Pi as the central node, RTL-SDR for signal sensing, and LoRa transceivers for communication. This dual demonstration validated the practicality and efficiency of the approach in addressing dynamic spectrum management challenges.