Research Publications & Projects

Seyed Moein Abtahi* and Akramul Azim

The International Conference on Predictive Models and Data Analytics in Software Engineering (PROMISE), 2026

This study introduces the Next Generation Quality Accelerator (NGQA) pipeline, a comprehensive framework that automates software quality assurance by intelligently orchestrating static analysis and large language models (LLMs). The NGQA methodology addresses the critical challenge of scalable quality assurance by establishing a systematic, six-step pipeline that progressively refines code quality while minimizing human intervention. The framework integrates SonarQube-based issue detection, retrieval-augmented generation (RAG)-driven false-positive mitigation, LLM-based code remediation, structural dependency analysis, comprehensive test suite generation via a novel Local Chain-of-Thought framework that employs four sequential AI agents, and multi-metric quality validation. Experimental evaluation across 70 repositories spanning seven programming languages demonstrates significant effectiveness: PassRatio improved by 16.5%, CodeBLEU increased by 28.8%, and CodeScore improved by 24.0%. The false-positive mitigation agent achieved an 89.0% F1-score, NGQA successfully resolved 83.5% of validated issues, and achieved a 32.6-fold mean acceleration over the estimated manual QA effort, representing a significant step toward largely automated software QA with minimal human intervention.

Muhammad Zaeem Khalid, Arhum Ahmed, Akramul Azim, and Taufiq Rahman

IEEE ITSC 2026

This work presents an energy-aware quadratic programming (QP) controller for connected and automated vehicle platoons that embeds a physics-based traction power model directly into the real-time control loop. Unlike standard adaptive cruise control (ACC) and cooperative ACC (CACC), which treat energy efficiency as a byproduct, the proposed framework jointly optimizes string stability and instantaneous power consumption at every time step. Evaluated in a CARLA-based co-simulation environment under sinusoidal leader disturbances for 3-vehicle and 5-vehicle platoons, the QP controller achieves up to a 23% reduction in energy consumption over standard ACC while maintaining the tightest variance and the best string stability ratios across all followers.

We are working on a study that addresses a gap in today’s autonomous systems: they record what happened, but not how an initial intent was refined into the actions the system finally took. Without that refinement trail, basic questions go unanswered like who defined the behavior, who enforced it, what the original intent was, how it changed along the way, and why the system acted as it did.

To address this, we introduce SPINE, a structured semantic policy-refinement framework where it treats the intent as a declarative policy and refines it step by step into definitive and finally imperative policies, preserving the meaning of the original intent throughout. We call these layers the abstraction hierarchy.

In addition, each policy carries a policy model made up of its definer, enforcer, action, constraints, and context, together with policy metadata (ID, domain, expiration, priority, and an autonomic-permission flag), so that every policy is fully traceable.

This structure gives SPINE four key properties:

• Semantic intent preservation: the intent’s meaning stays intact through refinement.
• Traceability: the full lineage from any policy back to the original intent can be recovered by following stored pointers.
• Accountability: who defined and who enforced each policy can be identified.
• Explanation fidelity: every enforcement can be explained directly from the policy that produced it.

In this study we focus our evaluation in two domains, code execution and autonomous vehicles, just to see whether it generates policies that match the intended semantics and whether it works as a general framework across very different kinds of intent. The goal is not heavy benchmarking but to introduce SPINE and show a practical foundation for building more structured, explainable autonomous systems.

The research explores how Vision-Language Models (VLMs), Vision-Language-Action (VLA) approaches, and deep reinforcement learning (DRL) can support autonomous vehicles in understanding complex driving scenes and translating that understanding into decisions, planning outputs, and actions.

The work investigates how visual and language information can be integrated to support vehicle decision making and macro-level planning. This includes a literature review, assessment of model architectures, training strategies, and evaluation methods, as well as experimental design and early prototyping where appropriate.

The broader goal is to identify feasible and effective approaches that can inform the development and evaluation of VLA and DRL based autonomous driving systems.

Mansi Sahi, Nitin Auluck, Akramul Azim, Pooja Bhardwaj, and Md. Al Maruf

IEEE Transactions on Network and Service Management, 2026

In this paper, we propose D2-TONE (Data-driven Deep Neural Network Task Offloading on the Network Edge), employing ML algorithms to accurately estimate offloading costs, such as computation and transmission costs. D2-TONE adapts holistically to dynamic network situations, and provides optimal/near-optimal offloading solutions in real-time. In addition, D2-TONE employs the distributed execution of DNN training tasks on the edge-cloud continuum. Experiments revealed that D2-TONE reduces the training time by 1.55 to 2.77 times, compared to baseline approaches. In addition, the edge based D2-TONE offers an improvement of 55-76% in the data processing ratio, compared to other offloading approaches.

Arhum Ahmed, Akramul Azim, and Taufiq Rahman

IEEE ITSC 2026

This study addresses the silent failure of deep learning models in predicting electric vehicle power consumption when real-world driving conditions deviate from training data due to weather, traffic, or sensor noise. The proposed framework combines a 30-network deep ensemble with distribution-free conformal prediction to generate prediction intervals with formal coverage guarantees, while using ensemble disagreement as a lightweight out-of-distribution detector. Tested across six realistic scenarios in CARLA, including snow, wind, rush hour, and combined perturbations, the approach maintains 97–100% coverage under severe distribution shifts, with a rolling conformal strategy tightening interval widths by approximately 37% compared to fixed calibration, making it particularly suitable for safety-critical energy management applications.

Ajwad Abrar and Akramul Azim

ICSE 2026 SEiGS

Software metrics are vital for ensuring software quality, yet their adoption and impact in fast-developing countries remain underexplored. This paper presents an empirical study of software metrics practices, tools, and challenges using Bangladesh as a representative example of a rapidly growing software industry. Through interviews with 20 professionals from 13 companies, we examine how metrics are applied across Software Development Life Cycle (SDLC) phases and identify major categories such as size, test, productivity, structural, and reliability metrics. Metrics such as test coverage, uptime, and coupling between objects are widely used to improve maintainability and error detection. Despite these benefits, adoption is limited by insufficient training, skepticism, and perceived workload. Common tools supporting metric-based development include Jira, ESLint, Postman, and Docker. The findings highlight how Bangladesh’s experience reflects broader trends in rapidly developing economies and provide insights for strengthening the adoption of software metrics globally.

Mohammad Abu Obaida Mullick, Rejuyan Ahmed, Md. Motaharul Islam, Muhammad Nazrul Islam, Akramul Azim, Bader Fahad Alkhamees, and Mohammad Mehedi Hassan

IEEE Access, October 2025

The transition from student life to professional employment represents a challenging shift marked by emotional and practical adjustments. This research explored the sentiment dynamics of two surveys conducted with two groups from the Science and Technology background: fresh graduates who have entered the workforce and students in the final phase of their academic life, likely to enter the workforce. Exploratory data analysis was performed to process numerical and categorical responses, and an autonomous thematic analysis was conducted after applying K-Means clustering to open-ended responses. Subsequently, Google Pegasus-Large was used to summarize open responses and assign preliminary clustering to the responses, and the GPT-5.1 model was utilized to extract themes and codes with sentiment labels, which were later reviewed through manual validation. By capturing and analyzing these sentiments, the research identified patterns in expectation management, resilience-building, and coping mechanisms for workplace stress. In this article, we provided guidance on targeted employee support and organizational adaptation, and supported human factors–informed interactive tools to optimize user experiences and ease the academic-to-work transition for new professionals. Finally, this research provided a foundation for improving employee well-being, effective organizational practices, and healthier work environments.

Amin Avan, Akramul Azim, and Qusay Mahmoud

35th IEEE International Conference on Collaborative Advances in Software and Computing (CASCON 2025)

Modern soft real-time applications (SRTAs) impose heavy computational demands on embedded devices. While offloading workloads to Edge Computing (EC) resources is attractive, task scheduling remains challenging due to strict timing constraints, a vast search space, multiple conflicting objectives, and highly dynamic environments. Conventional heuristic and meta-heuristic algorithms struggle to adapt to these conditions. Although reinforcement learning (RL) suits dynamic environments, single-agent RL converges slowly on medium- and large-scale problems due to enormous action spaces and excessive exploration. We present Agile Multi-Agent Reinforcement Learning (A-MARL), which enhances Multi-Agent PPO by replacing conventional exploration with entropy-guided rule-based exploration. When policy entropy is high, A-MARL employs Shortest Processing Time (SPT) to guide exploration toward promising action space regions. This adaptive mechanism accelerates convergence and delivers schedules better suited for SRTAs in EC environments. Experiments on representative scenarios show A-MARL consistently outperforms state-of-the-art baselines across all evaluated metrics, demonstrating its effectiveness for SRTA task scheduling in EC.

This main idea of this project is to integrate real hardware used in cars to connect with CARLA Simulator. This project develops a real-time driving environment and adds constraints to test different modules of autonomous vehicles. In addition, it aims to improve the software quality to bring more realistic and more modern features (e.g., Lane Assisting) for the users.