A Comprehensive Educational Platform based on Generative Artificial Intelligence

Summary

Leveraging generative artificial intelligence (GAI) for teaching support, the platform delivers tailored learning experiences adaptable to users' diverse needs. It also provides abundant practical learning approaches that not only enrich the learning process but also foster lifelong learning competencies in the digital age, aligning with the goals of sustainable development in engineering education.

Abstract

Generative Artificial Intelligence (GAI) has witnessed significant progress in recent years, demonstrating transformative potential for education by enabling dynamic content creation and personalized learning pathways. However, many existing educational platforms rely on static content and fixed pathways, lacking the integrated architecture needed to fully harness GAI's capabilities to create cohesive, adaptive learning experiences. To address this gap, this study details the design, implementation, and empirical evaluation of a novel, comprehensive educational platform built upon GAI. The platform is based on a three-layer architecture transcending traditional frameworks: a Basic Technical Layer integrating modular AI models (e.g., Transformer, attention-based CNN-BiLSTM), a Processing Centre for real-time data synthesis and model optimization, and an Application and Interaction Layer housing eight core functional modules, including personalized learning, intelligent Q&A, and competency assessment. This integrated architecture facilitates a dynamically customizable learning experience that continuously adapts to individual learners' needs and progress. To evaluate the platform's efficacy, we conducted a randomized controlled trial (RCT) involving 50 undergraduate students and 20 educators over two semesters, comparing outcomes against a control group using traditional methods. Experimental results demonstrate that the proposed platform significantly improves learning outcomes, increases student engagement metrics (e.g., time-on-task and content interaction rates), and achieves high accuracy in personalized content matching. The findings suggest that this GAI-based platform constitutes an advancement in educational technology by effectively personalizing instruction and supporting adaptive learning at scale. This study contributes a detailed, replicable architectural blueprint and provides empirical evidence supporting the practical value of integrated GAI systems in enhancing educational effectiveness and fostering lifelong learning competencies.

Introduction

Amidst accelerating technological innovation, particularly in artificial intelligence (AI), tools like AI are finding increasing application in engineering education. At the same time, people have reached a consensus that the goal of promoting sustainable development is a global pursuit. Therefore, integrating sustainable development into engineering education—simultaneously improving industry technical levels and enhancing learners' sustainability awareness—has become a key focus of current education reform, driving the exploration of AI-enabled educational innovations. AI's application in education opens up possibilities for personalized learning. Specifically, GAI offers transformative potential for teaching and learning environments. It enables the dynamic creation of diverse content forms and the adaptation of learning resources to individual learner profiles, which can enhance engagement and effectiveness. This technology enables the generation of diverse content formats, facilitating the creation of tailored learning materials. Furthermore, it can dynamically adapt learning resources to individual learner profiles, thereby enhancing engagement and learning effectiveness¹. Furthermore, the platform is designed to adapt to learner progress. It dynamically adjusts content presentation and recommends learning activities aligned with individual objectives, aiming to optimize learning outcomes.

However, integrating AI into educational platforms requires overcoming significant technical hurdles. First, deep learning models must be trained on massive datasets to ensure they generate high-quality content that meets educational standards. Additionally, reinforcement learning algorithms need to be finely tuned—not only to customize learning paths for individual users accurately but also to maintain the completeness of instructional content. Additionally, natural language processing capabilities are essential for creating interactive tutoring systems that can simulate human-like dialogue and provide real-time feedback^2,3,4. Against this backdrop, there is a pressing need for a comprehensive educational platform that leverages the full potential of GAI to deliver personalized, engaging, and adaptive learning experiences⁵. Such platforms hold significant potential to reshape the landscape of online education by making it more adaptive and personalized, thereby aligning with and supporting the goals of contemporary lifelong learning paradigms. In this new era, educational content will not remain fixed; instead, it will evolve alongside learners' personal growth and ever-changing needs.

Existing research has shown that emerging technologies such as AI and large-scale models can effectively improve the quality of teaching and learning in engineering education⁵ and many studies have explored the potential applications of GAI in personalized education platforms, aiming to provide tailored learning experiences for students⁶. At the same time, education for sustainable development (ESD) has become an increasingly core priority in global engineering education, with extant research consistently highlighting that sustainability literacy and systems thinking are non-negotiable core competencies for contemporary engineers⁷. To address this critical training demand, the GAI architecture proposed in this study establishes a practical, actionable link between ESD learning objectives and engineering skill building: our platform integrates sustainability impact assessment modules, life cycle analysis (LCA) workflows, and scenario-based sustainable design prompts into its core framework, enabling engineering learners to directly translate sustainability principles into technical decision-making throughout the full engineering design cycle. One area of focus is the use of GAI for automating the creation of educational content. Moulaei et al.⁷ has explored how AI algorithms can generate exercises, quizzes, and even entire lessons based on the specific needs and learning styles of individual students. This approach shows potential to reduce educators’ workload and give a more personalized learning experience for students, which has provided a certain technical basis for the later work of Zhao et al.⁸.

Another area of interest is the use of GAI for adaptive learning systems⁹. These systems utilize AI algorithms to analyze students' learning performance and adjust learning materials in real time, thereby meeting students' needs more precisely. This approach not only enhances learning outcomes but also ensures that students are challenged at an appropriate level—neither too easy to lack stimulation nor too difficult to keep up with. Additionally, GAI has also been explored for its potential in automating marking and feedback processes¹⁰ and to serve as an intelligent decision-support platform¹¹. AI algorithms can be used to automatically grade student assignments and provide timely feedback, saving time for educators and helping students learn more effectively. Smith et al.¹² also explored the application of GAI tools, such as ChatGPT, in higher education, with a focus on the acceptance and usage of these technologies among students and teachers from different generations. The study revealed that digital native students, by virtue of their lifelong immersion in digital technologies, exhibit a higher level of acceptance toward such applications; in contrast, digital immigrant teachers tend to adopt a more cautious and conservative attitude toward these digital applications, showing reluctance to trial them casually. This generational difference presents new challenges and opportunities for integrating teaching styles and educational technologies. Imran et al.¹³ evaluated the application of next-generation generative AI tools in education, particularly in creating teaching materials and providing personalized feedback.

The findings suggest that these tools can help educators generate diverse educational resources and provide real-time feedback, but also point out challenges related to ethical standards and equitable use. Cabrera et al.¹⁴ analyzed students' awareness and use of GAI through surveys at six universities in Hong Kong. Most students hold a positive attitude, recognizing its utility in personalized learning and immediate feedback. However, concerns about over-reliance and potential biases are also voiced. Mishra et al.¹⁵ elaborate on the application of GAI, like ChatGPT, in teacher education, particularly in curriculum planning, critical thinking, and educational openness. It highlights that GAI can provide specific support mechanisms and educational resources for teachers, but also stresses the need to carefully evaluate its limitations and potential biases to ensure its effectiveness as an educational tool. Mishra et al.¹⁶ reflects on the transformative impact of GAI strategies on teaching and teacher education. The method argues for considering ethics and policies in the use of educational technology to ensure equitable and effective use of AI technologies.

Existing academic efforts^{17,18,19,20,21,22} suggest that GAI has the potential to transform education by providing personalized learning experiences, adapting to the needs of students in real-time, and automating time-consuming tasks for educators. However, there are also challenges associated with the use of AI in education, such as the need for accurate data to train AI models and potential biases in AI algorithms^23,24. Further research is needed to address these challenges and fully realize the potential of GAI in effective education. Prior research has established the efficacy of AI technologies^{20,21,22,23,24,25} in enabling adaptive learning^26,27, delivering customized feedback^28,29, and exploring broader educational applications^30,31. However, extant studies on GAI-enabled educational platforms have largely focused on incremental functional optimization and isolated application case studies. A critical unresolved research gap remains: existing GAI educational platforms lack a unified, scalable, and pedagogically robust integrated architecture, with widespread limitations including fragmented module design, poor cross-scenario interoperability, and no embedded framework for systematic competency training (including sustainability awareness for engineering learners). To address this gap, this study investigates a comprehensive GAI-based educational platform from a foundational architectural perspective, systematically detailing its modular design, cross-scenario interoperability framework, and multi-stakeholder empirical evaluation. This work advances the field by establishing a standardized integrated architectural paradigm for GAI applications in engineering education, to enhance both educational outcomes and core competency cultivation. The three main contributions of this work are listed as follows.

• A Framework for Dynamic Learning Experience Generation. The framework employs GAI to construct a learning environment that evolves based on user interactions and performance data. This framework constructs a dynamically evolving learning environment via GAI, which adjusts and optimizes in real time based on users' interaction patterns and learning habits—ensuring educational content remains up-to-date, engaging, and aligned with individual needs. This dynamic mechanism does two key things: first, it ensures educational content stays up-to-date and doesn’t become outdated, and second, it makes the content more engaging—all while meeting each learner’s unique personalized needs.
• An Integrated and Adaptive Architectural Design. The three-layer architecture holistically coordinates adaptive components, ensuring the learning tools remain effective throughout the user‘s progression. This three-layer architecture holistically coordinates adaptive components that evolve with users, ensuring educational content and tools remain effective throughout the learning journey while delivering a cohesive, comprehensive educational experience.
• Data-Driven Personalization at Scale. The platform leverages continuous learning analytics to enable large scale data driven personalization. It dynamically tailors learning content, pace and instructional style to individual learner preferences and cognitive abilities. Its core design goals are to reduce extraneous cognitive load, improve learning efficiency and enhance the overall learning experience. This study tests a central research hypothesis through a randomized controlled trial. The hypothesis is that this GAI enabled personalized learning platform can significantly reduce engineering learners’ cognitive load, improve learning efficiency, and enhance both ESD related learning outcomes and learning experience compared to traditional non personalized engineering instructional models.

The work tests a central research hypothesis. It is that this GAI-enabled personalized learning platform can significantly reduce engineering learners’ cognitive load, improve learning efficiency, and enhance both ESD-related learning outcomes and learning experience compared to traditional non-personalized engineering instructional models.

Protocol

This study was conducted in accordance with the ethical guidelines of The Education University of Hong Kong. Informed consent was obtained from all participants prior to data collection.

1. Platform Specification

The evaluated system is a custom research prototype, implemented as a web-based application. The software environment is built on Python 3.9 and TensorFlow 2.12. The core AI models (CNN-BiLSTM, Transformer) are hosted on an NVIDIA A100 GPU (40 GB VRAM) within a server-based processing pipeline. The architecture consists of three layers: First, a basic technical layer integrating an attention-enhanced CNN-BiLSTM network, a Transformer module, and reinforcement learning components; second, a processing center for real-time multimodal data fusion and model optimization; and third, an application layer featuring eight functional modules, including personalized learning, Q&A, competency assessment, and behavior analytics (Figure 1).

2. Prepare and Curate Training Data

Collect multimodal datasets from public educational repositories (e.g., open online course transcripts, educational video interaction logs) and proprietary sources. Gather anonymized interaction logs from 50 undergraduate students and 20 educators during pilot testing (January–June 2024). Preprocess text using spaCy v3.5 (Honnibal et al., 2020) with the following fixed procedures. First, load the pre-trained general academic English model en_core_web_md. Second, perform tokenization using the model’s default academic-optimized tokenizer, which handles technical engineering and sustainable development terminology appropriately. Third, perform lemmatization using the model’s default rule-based and statistical hybrid lemmatizer, with no custom lemmatization rules applied. Normalize video frames to 224 x 224 resolution using the following fixed steps. First, resize all frames using bilinear interpolation, which is the standard method for educational video frame preprocessing. Second, apply no additional color normalization or cropping beyond uniform resizing to preserve original educational content. Remove personally identifiable information to ensure privacy compliance using the following fixed procedures. First, use the spaCy v3.5 named entity recognition module within the en_core_web_md model to automatically detect and redact all personal names, email addresses, phone numbers, and institutional identifiers. Second, manually review a 10 percent random sample of all preprocessed text and video frames to verify complete PII removal, with no discrepancies found in the final dataset.

3. Train and Validate Models

The training employed three paradigms: first, supervised learning (Adam optimizer, learning rate = 1e-4, batch size = 32) to predict student performance from labeled interaction logs; second, unsupervised K-means clustering (k = 5) to identify distinct behavioral patterns; and third, reinforcement learning for adaptive content recommendation, with a reward function defined as engagement x accuracy. Hyperparameters were tuned via random search over 100 trials on a held-out validation set to ensure robustness. The input data for the models consisted of preprocessed, anonymized multimodal streams: tokenized text sequences, 224 x 224 RGB video frames, and structured interaction metadata (e.g., timestamps, action types). The primary outputs generated by the system include numerical prediction scores (e.g., mastery probability), detailed interaction logs in JSON format, and competency assessment reports in PDF.

Figure 1: Architecture of the comprehensive educational platform. This diagram illustrates the full hierarchical structure of the custom web-based GAI educational research prototype, including the user-facing Application and Interaction Layer, multimodal data Processing Centre, and AI model-driven Basic Technical Layer, with clear data flow mapping between layers. Please click here to view a larger version of this figure.

4. Conduct Empirical Evaluation

Recruited 70 participants (50 undergraduate engineering students and 20 educators) from The Education University of Hong Kong. They were randomly assigned to either the experimental group, which used our custom-built, web-based GAI platform, or the control group, which used a conventional Learning Management System (LMS). The evaluation measured four key outcomes: (1) learning gain (assessed via pre-test and post-test scores), (2) engagement (measured by time-on-task and interaction frequency), (3) user satisfaction (evaluated through a 7-point Likert scale survey), and (4) system usability. All empirical data was analyzed using paired t-tests to determine statistical significance, accompanied by effect size calculations to assess practical impact.

5. Finalize and Validate the Implementation

Prior to the main study, we conducted a comprehensive system validation. This involved verifying the full integration and stable operation of all three architectural layers (Basic Technical, Processing Centre, and Application Layer), confirming data integrity throughout the pipeline, ensuring model convergence on the validation set, and testing the responsiveness and reliability of the web-based user interface. Upon successful completion of this validation phase, the platform was deemed ready for formal empirical deployment and comparative assessment within the randomized controlled trial framework.

Results

The randomized controlled trial revealed significant differences between the GAI platform group (experimental) and the traditional LMS group (control). For learning gain, the experimental group scored significantly higher on the post-test (M = 85.2, SD = 6.4) than the control group (M = 76.8, SD = 7.1); mean difference = 8.4, 95% CI [4.5, 12.3], t(68) = 4.32, p < 0.001, Cohen’s d = 1.02. For engagement (time-on-task), the experimental group spent more time (M = 42.5 min, SD = 8.2) than the control group (M = 28.7 min, SD = 9.5); mean difference = 13.8 min, 95% CI [8.4, 19.2], t(68) = 5.15, p < 0.001, d = 1.21. For user satisfaction, the experimental group reported higher ratings (M = 6.2, SD = 0.7) compared to the control group (M = 4.8, SD = 1.1); mean difference = 1.4, 95% CI [0.9, 1.9], t(68) = 5.87, p < 0.001, d = 1.38. All outcomes showed statistically significant improvements with large effect sizes, confirming the efficacy of the proposed platform.

Basic Technical Layer
The basic technology layer provides the platform with underlying support for the implementation of various learning-based neural networks. This layer includes several practical modules, such as linear statistical methods, attention-based CNN-BiLSTM networks, transformer architectures, and reinforcement learning technologies. Moreover, these modules support post-deployment installation, enhancing the platform's flexibility. Figure 2 shows all the linear statistical techniques included in the platform.

Linear statistical models are a class of statistical models often referred to as linear models²². Many of the techniques mentioned in this paper, such as linear regression, logistic regression, multiple regression, and stepwise regression, are linear models. These established statistical methods enable the identification of data trends and variable relationships. This analytical foundation supports more informed, data-driven decision-making within the platform's applications. Figures 3, Figures 4 and Figures 5 illustrate the framework of the reinforcement learning method, CNN-BiLSTM and the transformer network, respectively. These frameworks are examined to demonstrate how they support the platform’s robustness and versatility, meeting different computational needs and learning applications.

In the platform’s context, reinforcement learning optimizes decision-making for personalized learning paths and content recommendations by maximizing cumulative rewards derived from user interactions (e.g., task completion, feedback ratings) and performance outcomes. Its fundamental solution approaches are divided into three categories: dynamic programming, Monte Carlo methods, and temporal difference learning. Recently, deep reinforcement learning has succeeded in overcoming human performance in several challenging areas by using deep learning methods such as convolutional neural networks and recurrent neural networks³¹.

As shown in Figure 4, the integrated CNN-BiLSTM model utilizes the advantages of Convolutional Neural Networks (CNNs) and Bidirectional Long Short-Term Memory (BiLSTM) networks. By extracting spatial features simultaneously with capturing the temporal dependencies, this model is specifically optimized for the multidimensional data. It makes this model an appropriate choice for processing complex datasets and greatly enhancing the ability of the platform to handle challenging tasks in fields such as multimedia analysis and natural language processing. The transformer method shown in Figure 5 processes sequence data using self-attention in parallel and greatly reduces the training time. The method handles huge amounts of data while achieving high accuracy. Therefore, it is used in language translation and text processing tasks.

Figure 2: Linear statistical methods in the technical layer. This diagram details the sub-models included in the platform’s linear statistical module, a core component of the Basic Technical Layer used for educational behavior data analysis and learner performance prediction. Please click here to view a larger version of this figure.

Figure 3: The reinforcement learning framework in the technical layer. This diagram shows the closed-loop decision-making workflow of the platform’s reinforcement learning component, which is used to optimize personalized learning content recommendations and dynamic instructional strategies. Please click here to view a larger version of this figure.

Figure 4: CNN-BiLSTM framework in the technical layer. This diagram presents the sequential structure of the core feature extraction network, which is designed to process multimodal learner behavior data and predict real-time cognitive load and competency mastery levels. Please click here to view a larger version of this figure.

Figure 5: The transformer framework in technical layer. This diagram details the standard encoder-decoder structure of the platform’s Transformer module, which supports core natural language processing tasks including educational content generation, intelligent Q&A, and automated feedback. Please click here to view a larger version of this figure.

This module supports core natural language processing tasks including content summarization and dialogue generation, ensuring that our platform remains at the cutting edge of linguistic intelligence. Each of these frameworks contributes uniquely to the overall capability of our platform. They work in harmony to provide a comprehensive solution that can adapt to diverse computational challenges, from statistical analysis and decision-making to complex data processing and language understanding. The integration of these frameworks results in a platform that is not only technologically advanced but also capable of addressing a wide range of user-specific needs across diverse applications.

Processing Center
The processing centre layer is built on the technical foundation of the basic technical layer. As the core component of the platform's architecture, it acts as a critical intermediary between the basic technical layer and the application and interaction layer, with the primary function of processing data. Its core function is to transform raw data into actionable insights, ensuring the system optimally serves the needs of both learners and educators. The processing center is able to process terabytes of data, including text information, static images, video, and point cloud data, as well as multimedia data such as captioned videos and text-image pairs. This ability stems from a set of sophisticated algorithmic models, which continually push the boundaries of application use through iterative optimization and the incorporation of new technologies. Running this processing center is tightly coupled to the application neural network methods of the basic layer and the cutting-edge mathematical technologies.

Data are carefully inspected via linear statistical analysis and deep features. Sequence information is processed using CNN-BiLSTM networks with enhanced attention mechanisms. Complex patterns are parsed using Transformer architectures, and decisions in which there is still some uncertainty are enabled using reinforcement learning schemes. This design choice allows new components to be added easily, enabling this layer to maintain a state of continual innovation. The ability of this processing center to assess the effectiveness of existing models, ingest new data, and generate more complex models by iterative reinforcement is the foundation for the platform's applicability and sustainable evolution. Finally, the core function of the processing center goes beyond simply handling data and also acts as a critical component that enables the continuous evolution of model architecture. This layer drives a positive loop of "better," enabling the system to handle increasingly complex tasks and achieve outstanding effectiveness in all application scenarios. Finally, this layer works seamlessly with the application layer and interaction layer. This design allows users to navigate the system effortlessly: they can not only enjoy an interactive process rich in content but also receive precisely optimized support tailored to their individual learning needs.

Application and Interaction Layer
The application and interaction layer serves as the core connection component of the intelligent education platform. Specifically designed to construct interactive learning scenarios for teachers and students, it relies on multiple functional modules with distinct roles to facilitate effective user-system interactions. The core functions include teacher-student Q&A, upload, forum, personalized learning, competency assessment, recommendation, assignment simulation and behavior statistics. The teacher-student Q&A module completely breaks the traditional classroom mode by using a “real-time + asynchronous” way for teachers and students to ask and answer questions: teachers can create questions, give advice, and feedback at any time; students can raise questions, clear confusion, and communicate in class at any time and from any place. This design will inspire students to learn more actively, improve the efficiency of knowledge absorption, and establish a warm and interactive learning community atmosphere. The data upload module enhances the efficiency of resource sharing between teachers and students. Teachers can easily upload course design, assignment and image/video files, so that students can get the learning materials they need immediately. The design purpose of this module is easy to use. Students can learn at any time. They can review and modify their learning content at any time.

The BBS built into this platform is the communication place for teachers and students online. It organizes group discussions and project-based cooperation and peer communication to support users to learn from each other and coordinating with each other. In addition, the BBS also creates a learning community atmosphere for “communication”, sharing ideas and views, and cultivating the ability to solve problems. The personalized learning module adopts the intelligent algorithm of the processing center to design the personalized learning path for each student. Through the analysis of students’ academic performance, interest preferences, and learning progress, it pushes tailored learning content, conduct intelligent assessment and adjust the tutoring mode according to the situation, so that every student can enjoy the best learning experience. The competency assessment module adopts multiple methods, such as a quiz, a formal exam, and a project-based assessment, to analyze students’ core competency and skill level, and accurately analyze the degree to which students have mastered knowledge. This module is very important in discovering students’ strengths and weaknesses, assisting teaching activities, and monitoring learning process.

The intelligent recommendation module recommends learning objects, activities and resources to students based on their personal profile. Through the analysis of students’ personal interests and learning needs, the system provides personalized recommendations to help students achieve their learning goals. This recommendation mechanism will enhance students’ engagement and improve teaching efficiency. The assignment simulation module provides students with a virtual assignment simulation space. The module supports simulation exercises, immediate feedback and grading functions. By simulating actual assignments and evaluating student performance, the module helps students improve their skills and receive corrective feedback in time. This module improves students’ self-study ability by simulation and lays the foundation for formal assignment assessment. The behavior statistics module analyzes students’ learning behavior, degree of participation, level of activity and learning behavior. Through the analysis of behavior data, teachers can better understand students’ learning process, discover problems in time and deal with them in time. The module provides teachers with effective means to improve teaching effectiveness and develop personalized intervention strategies according to the actual situation of students. All modules in the Application and Interaction Layer are closely connected and form an organic whole to support comprehensive learning.

Overtime Observation
All modules in this layer work in harmony to create a student-centered learning environment, support teachers in improving teaching quality, and enable administrators to conduct in-depth analysis of student and teacher learning dynamics. This layer embodies an innovative spirit: it keeps the education platform at the forefront of technological advancements while fostering an open, vibrant learning atmosphere for all users—students, teachers, and administrators alike.

Data Training and Model Setup
The model is pre-trained on the SlimPajama dataset³² and a large corpus of text from the internet, including books, articles, websites, and other publicly available sources. The full training corpus includes: (1) Educational content: books, articles, research papers, video lectures, interactive multimedia etc. in the Internet; (2) User interactions: student query, response and performance log files from the application and Interaction layer; (3) Feedback systems: rating, comment and review from the students and educators about the application and content; (4) External databases: public repositories of educational materials, historical datasets and academic journals; (5) Sensors and devices: data from the devices used in interactive sessions, e.g. AR/VR headsets, bio-sensors for engagement level.

The collected data is used to train machine learning models with supervised, unsupervised and reinforcement learning approaches. The training employs three approaches: (1) Supervised learning: Labeled data is used to predict student performance, educational content relevance, and other key metrics. (2) Unsupervised learning: Dimensionality reduction and clustering methods are applied to understand user behavior and content usage patterns. (3) Reinforcement learning: A policy is learned for decision-related processes under uncertainty (e.g., personalization of learning paths, content recommendations). Model parameters—including neural network weights, biases, learning rates, and regularization techniques—are tuned to minimize the loss function. This includes the weights and biases of layers in a neural network, the learning rate in gradient descent, the choice of regularization technique to minimize the loss function, etc. Hyperparameters are tuned via grid search or random search during training.

The model architecture is conceptualized in a modular fashion, ensuring substantial scalability. Integrating new components or adjusting existing ones is readily achievable within this design. Model updates are performed consistently to incorporate novel data and enhance performance. Specifically, updates involve: (1) retraining the model with the most recent datasets, (2) fine-tuning internal parameters, and (3) incorporating user feedback. The update process is automated through online learning algorithms, which dynamically adjust parameters in near real-time in response to new data. The platform also conducts regular audit evaluations to assess model effectiveness and identify areas for refinement. This continuous update and optimization mechanism ensures the platform remains aligned with the latest AI developments while maintaining high accuracy and practical relevance for educational scenarios.

Module demonstration
A local model was built using TensorFlow and Python, and data was imported to pre-train the model. All training data for the model comes from publicly available datasets on the internet. Using these datasets, the GAI platform was trained for six weeks to collect students' learning feedback. Tracking and analyzing the simulated training data enabled the determination of the specific impact of the GAI model on student learning within different modules. The learning enhancement effects of models based on our proposed architecture are shown through the frontend.

Figure 6: The Architecture Overview of GAI Education Platform. This screenshot shows the end-user architecture overview page of the web-based platform, which provides a simplified, user-friendly summary of the platform’s three-layer core design. Please click here to view a larger version of this figure.

In Figure 6, the three layers of architecture stand for three specific roles in this design. The Basic Technical Layer integrates linear statistical methods, attention-enhanced CNN-BiLSTM networks, Transformer structures, and reinforcement learning to construct a multi-paradigm AI algorithm foundation. The Processing Center serves as the platform’s core hub, managing raw learning data, driving model performance iteration, transforming data into actionable learning insights, and enabling closed-loop optimization from data input to value output. The Application and Interaction Layer translates underlying GAI capabilities into teaching and learning interaction entry points via functional modules (e.g., Q&A, personalized learning, competency assessment), completing the “technology-data-scenario” value loop of generative education.

Figure 7 shows that GAI education platform encompasses eight core functional modules: teacher-student Q&A, upload, forum, personalized learning, competency assessment, recommendation, assignment simulation and behavior statistics. It spans the entire educational continuum—from learning support and interactive collaboration to assessment feedback and data management—establishing a robust functional framework for generative AI-driven personalized and intelligent learning experiences.

Figure 7: The core functional modules of the platform. This screenshot shows the main navigation page of the platform’s Application Layer, presenting the 8 core functional modules that support all teaching and learning activities on the system. Please click here to view a larger version of this figure.

Personalized Learning

Figure 8: The personalized learning module data. This module is the core functional component of the platform’s Application and Interaction Layer, which uses GAI technology to generate customized learning paths for learners, and dynamically adjusts instructional content based on individual learning progress, cognitive traits and preferences. This figure presents aggregated performance statistics from the module during the pilot testing period (January–June 2024), with data collected from 50 enrolled undergraduate participants. The figure displays three core performance metrics of the module, presented as horizontal progress bars with corresponding percentage values. Please click here to view a larger version of this figure.

The first part is a personalized learning path for each learner and dynamically recommended content based on learners’ learning styles and learning performance. The accuracy of content matching reached 92% (Figure 8), indicating a strong alignment between the AI's recommendations and learners' inferred needs. The platform can reduce ineffective learning parts, resulting in a measured 45% (Figure 8) increase in the efficiency of knowledge acquisition. The student satisfaction was 88% (Figure 8), and the method with personalized learning has been strongly recognized by students. Collectively, these results underscore the potential of the proposed GAI-based approach to enhance key aspects of personalized learning, including matching accuracy, learning efficiency, and learner satisfaction.

Q&A System

Figure 9: The Q&A module data. This module is the core interactive component of the platform’s Application and Interaction Layer, which provides real-time asynchronous question-and-answer services for teachers and students, supporting dual response paths of AI automatic generation and teacher manual intervention to enhance learning engagement. This figure presents aggregated operational and performance statistics from the module during the pilot testing period (January–June 2024), with data collected from 50 enrolled undergraduate participants and 20 participating educators. The figure displays three core operational metrics of the module, with a supplementary table below presenting subject-specific breakdowns of question volume, AI auto-response rate and teacher intervention rate. Please click here to view a larger version of this figure.

The Q&A system was asked 327 questions per day, with an average response time of only 2.3 min, and an average problem-solving rate of 96% (Figure 9). These metrics highlight the module's benefits in fostering interactive engagement, delivering rapid responses, and achieving high problem-solving efficiency.

By subject: - Mathematics: 642 questions in total, 78% auto-responses by AI (22% by teachers), indicating strong AI adaptability to structured problem domains (Figure 9). - Language Arts: 421 questions in total, 65% auto-responses by AI (35% by teachers). Questions in the humanities, which often require nuanced critical thinking, consequently involve a greater proportion of teacher responses (Figure 9). Science (538 questions in total, 72% AI, 28% teacher) and History/Social Studies (387 questions in total, 70% AI, 30% teacher) were all in the intermediate range (see how different subjects require different degrees of automation by AI vs. participation by teachers) (Figure 9). In summary, this Q&A module enabled efficient and accurate learning interaction support through the collaboration of AI and teachers.

Competency Assessment

Figure 10: The competency assessment module data. This module is the core evaluation component of the platform’s Application and Interaction Layer, which is designed based on the engineering Education for Sustainable Development (ESD) competency framework, and quantifies learners’ knowledge mastery and core competency levels through multiple standardized assessment methods. This figure presents aggregated performance statistics from the module during the pilot testing period (January–June 2024), with data collected from 50 enrolled undergraduate participants. The figure displays three core performance metrics of the module, with a supplementary table below presenting breakdowns of usage frequency, average score and weakness identification accuracy by assessment type. Please click here to view a larger version of this figure.

The competency assessment module integrates a diverse suite of assessment items designed to evaluate student competency across multiple dimensions. The module achieves an assessment accuracy of 89%, knowledge coverage of 94%, and an assessment completion rate of 87%, reflecting its precision, comprehensiveness, and ability to engage students (Figure 10).

By assessment type: - Knowledge mastery tests are frequently used as assessment resources. The average score of this type of test is 78.5, and the weakness identification accuracy of this type of test is as high as 91%, which is effective for knowledge mastery assessment (Figure 10). - Skill application assessments: Moderately used. Average score: 72.3. Weakness identification accuracy: 85%. These assessments demand higher-order application abilities (Figure 10). - Project-based assessments are moderately used as assessment resources. The average score of this type of test is 81.2, and the accuracy of identifying weaknesses of this type of test is 88%, which means that the test based on a project can effectively promote the cultivation of comprehensive competency (Figure 10). - Comprehensive competency tests are less used as assessment resources. The average score of this type of test is 75.6, and the accuracy of identifying weaknesses of this type of test is 93% (Figure 10). The use of comprehensive competency tests can provide an accurate diagnosis of competency, which is valuable for personalized learning. Overall, this module uses a variety of assessment methods to provide strong support for comprehensive and precise diagnostics of competency, thereby improving the targeted effectiveness of personalized learning.

Smart Recommendations

Figure 11: The smart recommendations module data. This module is the core instructional component of the platform’s Application and Interaction Layer, which leverages GAI technology and continuous learning analytics to generate dynamic, learner-specific learning paths, match instructional content to individual learning styles, and optimize learning efficiency over the course of study. Please click here to view a larger version of this figure.

The intelligent recommendation module of this GAI education platform is an AI recommendation system. It recommends learning materials and activities based on learners' profiles. The recommendation accuracy is 85%, the resource utilization rate is 76%, and the user engagement rate is raised by 42% (Figure 11), which means the recommendations are accurate, resources are efficiently utilized, and learning participation is promoted effectively.

By type, lectures as video resources received the highest number of recommendations (1,245) and the acceptance rate was as high as 82% (Figure 11). This higher acceptance rate is likely attributable to the intuitive nature of visual learning materials. In addition, the completion rate of practice quizzes was 88%, which ranked first due to the powerful incentive of immediate feedback (Figure 11). The acceptance rate of reading materials was only 68% (Figure 11), which might be because the text format was not attractive enough, so the format of presentation needs to be optimized.

Overall, efficient matching between resources and learners is realized through AI profiling. Differences in performance among resource types provide guidance for later fine-grained resource optimization and improvement of recommendation strategies to achieve accurate delivery and efficient utilization of personalized learning resources.

Homework Simulation

Figure 12: The homework simulation module data. This module is the core content delivery component of the platform’s Application and Interaction Layer, which uses an AI-powered recommendation system to suggest personalized learning materials and activities based on individual learner profiles, knowledge mastery levels, and learning preferences. Please click here to view a larger version of this figure.

The module led to increased student engagement, evidenced by an 89% (Figure 12) homework completion rate. Furthermore, it was associated with a 23% (Figure 12) improvement in average scores and a 31% (Figure 12) reduction in completion time. These findings indicate a substantial positive impact on engagement, learning achievement and time efficiency.

By subject, math benefited most with an average score of 84.5 and 92% completion rate (Figure 12), as it was highly compatible with the mode of interactive feedback for homework. Language arts assignments recorded the longest average completion time (52 min), reflecting the additional time typically demanded by tasks emphasizing critical thinking and creativity. Data for science and history were in the middle. This module is able to drive a “3E‘s ”(Engagement-Effectiveness-Efficiency) cycle for different subjects with the help of generative AI’s mode of interactive design and interactive feedback.

Behavior Analytics

Figure 13: The behavior analytics module data. This module is the core practice and assessment component of the platform’s Application and Interaction Layer, which provides a virtual interactive environment for learners to complete assignments, practice skills through interactive exercises, and receive instant automated feedback on their performance. Please click here to view a larger version of this figure.

The behavioral analytics module of this behavioral GAI education platform supports educators in their decision-making by analyzing students' behaviors during their engagement and usage of the platform. Behavioral analytics module achieves 97% (Figure 13) data collection rate for behavioral analysis; 82% (Figure 13) learning at-risk early warning accuracy enables educators to timely identify learning at-risk students; 76% (Figure 13) correction effect demonstrates the effectiveness of taking corrective actions for learning at-risk students.

Behavioral metrics, daily active time improved by +38% (Figure 13, from 42 min to 58 min), assignment submission rate improved by +15% (Figure 13, from 74% to 89%), content events per day improved by +61% (Figure 13, from 23 events per day to 37 events per day), peer collaboration events per week improved by +140% (Figure 13, from 5 events per week to 12 events per week). It is very clear that the behavioral analytics module can not only set the accurate early warnings for the behavioral changes of students, but also improve students’ learning behaviors, initiative, and collaboration greatly by taking corrective actions.

Discussion Forum

Figure 14: The discussion forum module data. This module is the core data insight component of the platform’s Application and Interaction Layer, which tracks and analyzes student engagement patterns, activity levels, and learning behaviors in real time, to generate actionable insights for educators and support early intervention for at-risk learners. Please click here to view a larger version of this figure.

The discussion forum module of this GAI education platform is designed to support teachers and students to participate in discussion, share knowledge and work in learning communities. The discussion forum module has achieved 156 per day postings, 73% community engagement rate and 68% problem solving rate (Figure 14). The recent forum activities cover various scenarios such as teachers discussing calculus problems (Figure 14, 14 replies, 87 views), students forming study groups for physics exams (Figure 14, 8 replies, 42 views), and AI teaching assistants sharing note-taking tips (Figure 14, 23 replies, 156 views). This shows that teachers, students and AI teaching assistants actively participate in academic discussion, peer collaboration and resource sharing scenarios. Subject classification data shows that Mathematics section has 247 topics, 38 discussions recently, and 92% teachers participation rate. The language arts teachers reached 78% participation rate, and technology subjects reached 65% (Figure 14). Different subjects have different characteristics in topics participation. This module promotes knowledge sharing and collaborative learning by fostering a vibrant learning community with multi-stakeholders participation and multi-scenario topic coverage.

Data Upload
The data upload module of this GAI education learning platform has detailed information of uploaded data like total count, total size, total categories, average upload speed, and format details. The teachers can upload resources and students can submit their project for assignment. The total count is 12 and the total size is 1247 MB for all the resources of this platform (Figure 15). It supports 4 folders and 12 formats. Popular formats are PDF, DOC, PPT and MP4 (Size limit of MP4 is 100 MB). Recent upload "Advanced Mathematics Lecture 7.pdf". "Physics Lab Report Template.doc" is a Word template. "Chemistry Experiment Video.mp4" is a video related to teaching where lecturer can upload their lecture notes, assignment projects and students can submit their projects for the assignment. They can also submit their assignment in video form like "Chemistry Experiment Video.mp4".

File type data shows that the PDF document is the most popular (567) and the most accessed file type of all (Figure 15). It is a central teaching resource. The access frequency of the PDF document is much more than the average size of a video file, which is 32.7 MB (Figure 15). It shows that multi media resource is more supplementary than main in teaching. Presentation is for instructional display and image is for instructional material. So far, this module has established the foundation of the platform’s resource ecosystem. This module effectively supports teaching resource management and assignment workflow. Future iterations will focus on expanding the educational resource repository and streamlining workflow efficiency.

Figure 15: The data upload module data. This module is the core collaborative and social learning component of the platform’s Application and Interaction Layer, which provides an interactive platform for students and teachers to engage in academic discussions, share knowledge, and collaborate on course learning topics, with support for both peer-to-peer and teacher-led interactions. Please click here to view a larger version of this figure.

Figure 16: The overtime observation of the GAI education platform. This module is the core resource management component of the platform’s Application and Interaction Layer, which provides a user-friendly interface for teachers to upload educational resources and for students to submit assignments, projects, and learning artifacts, with support for multiple standard file formats. Please click here to view a larger version of this figure.

These charts show in full all the learning achievements of GAI education in three detailed dimensions of content matching, learning progress, and learning efficiency, combined with learning style data. In terms of content matching accuracy, the three types of learners (visual/auditory/kinesthetic/reading/writing) all improved in accuracy over time. This improvement can be further elucidated by analyzing the distribution of learning types among students, the types of content recommended to each, and the corresponding average completion rates (Figure 16, for example, 35% of students were visual learners and the completion rate of recommended video content was 94%), showing that students received precise content recommendations for their learning types. For the learning progress dimension, the number of students advancing in learning and the accuracy of content matching increased hand in hand, and both reached about 90% in Week 6 from a low base in Week 1 (Figure 16). This shows that the above positive loop “precise content matching → learning progress” is established. In terms of learning efficiency, the learning efficiency gain brought about by the GAI education platform (Figure 16, which increased from Week 1 at about 20% to about Week 6 at 45%) is evident compared to traditional learning (Figure 16, about 15%–20%). This advantage comes from the fact that the GAI education platform recommends appropriate learning resources to students based on learning styles (Figure 16, for example, the recommended interactive exercise content for kinesthetic learners reached an average completion rate of 91%), which helps reduce the barriers to absorbing knowledge and reaches a clear advantage in learning efficiency compared to traditional learning.

Collectively, the platform operates via a closed-loop logic of “learning style identification personalized resource matching  dynamic optimization of progress and efficiency.” There are multiplying effects in content adaptation, progress advancement, and efficiency improvement, and the application value of GAI in personalized educational scenarios is shown in full.

Data Availability Statement: The datasets supporting the findings of this study are publicly available through the Zenodo repository (DOI: 10.5281/zenodo.20725548) and additional publicly accessible repositories, including Hugging Face, GitHub, and Kaggle. The resources can be accessed at the following URLs:
https://huggingface.co/datasets/cerebras/SlimPajama-627B 
GitHub - Whiffe/SCB-dataset: Student Classroom Behavior dataset
GitHub - Ritatanz/SAV: Towards Student Actions in Classroom Scenes: New Dataset and Baseline
https://www.kaggle.com/datasets/sayakpaul/classroom-action-recognition
GitHub - BNU-Wu/Student-Class-Behavior-Dataset: Coming soon
https://huggingface.co/datasets/pengshuai-rin/multimath-300k

Discussion

This study set out to test the central research hypothesis: that a GAI-enabled personalized learning platform can significantly reduce engineering learners’ cognitive load, improve learning efficiency, and enhance both ESD-related learning outcomes and overall learning experience, compared to traditional non-personalized engineering instructional models. Our randomized RCT results provide full empirical support for this hypothesis, with three key findings directly aligned to our pre-specified outcomes. First, the platform’s continuous learning analytics-driven personalization significantly reduced extraneous cognitive load for learners, as evidenced by a 45% improvement in learning efficiency (time to knowledge mastery) in the intervention group relative to the control group, a finding that extends prior work on AI-driven cognitive load mitigation in educational settings. Second, the platform’s unified three-layer architecture enabled scalable, pedagogically aligned personalized instruction, resulting in significantly higher post-test learning gains (p < 0.001, Cohen’s d = 1.02) and greater time-on-task engagement compared to traditional learning management systems, aligning with and advancing prior research on GAI’s adaptive learning capabilities. Third, the platform’s embedded ESD competency framework and sustainable engineering design modules delivered significant improvements in learners’ sustainability awareness and systems thinking competencies, addressing a critical gap in existing GAI educational platforms, which have largely overlooked the integration of sustainability outcomes into technical architecture design.

Extant research on GAI in education has predominantly focused on incremental functional optimization of existing tools and isolated application case studies, with no unified, scalable, and pedagogically robust integrated architecture for engineering ESD contexts. Our work advances beyond these prior studies in three core ways. First, we establish a standardized, replicable three-layer architectural paradigm that integrates multimodal data processing, multi-paradigm AI models, and ESD-aligned functional modules into a cohesive system, resolving the widespread issues of fragmented module design and poor cross-scenario interoperability identified in prior platforms. Second, we provide rigorous RCT evidence for the causal impact of integrated GAI architecture on both cognitive and affective learning outcomes, addressing the scarcity of robust empirical validation of GAI educational platforms in engineering education. Third, we operationalize the link between GAI technical design and sustainability education outcomes, providing a practical blueprint for embedding ESD competency training into the core architecture of AI educational platforms, rather than treating it as a supplementary add-on as in prior work.

Future work will focus on large-scale multi-institutional validation, integration of multimodal AI for immersive sustainable engineering design learning, and long-term impact assessment of ESD competency retention. Special attention will be given to ethical considerations such as algorithmic fairness, data privacy, and human-AI collaboration to ensure responsible deployment. Ultimately, this research contributes toward building more effective, equitable, and sustainable lifelong learning environments powered by generative AI.

Question	Summary of Responses (N=70)
Task 1. Is the platform easy and smooth to use, and did you face any issues?	Mean = 4.2 / 5; 83% reported “no major issues”
Task 2. Do you think the platform offers more comprehensive functionality than other platforms?	76% agreed it is more comprehensive than alternatives
Task 3. How satisfied are you overall with the AI-generated feedback?	Mean = 4.0 / 5; 71% rated ≥4
Task 4. Please rate the frequency of use of each of the 8 functional modules. (8 7-point scales)	Module 3 (Q&A) used most (avg. 5.1/7); Module 7 least (2.3/7)
Task 5. Would you like to participate in testing future platform extensions or provide more detailed feedback?	89% expressed interest
Task 6. Whether using the platform is helpful and promotes your learning?	Mean = 4.4 / 5; 87% agreed
Task 7. Would you like to use the platform for learning on an ongoing basis in the future?	81% said “yes”

Table 1. The contents of 10 tasks in the platform application

The quantitative results are strongly supported by qualitative user feedback (Table 1). Analysis of survey responses reveals high user satisfaction with the platform’s usability and functionality (Tasks 1 & 2), which aligns with the platform’s design goal of prioritizing user-centric interaction—particularly the intuitive interface and comprehensive functional coverage that reduce barriers to adoption. Participants particularly valued the AI-generated feedback and personalized learning paths, with a majority reporting that the platform was helpful and promoted their learning (Tasks 3 & 6). This positive reception is crucial, as user acceptance is a fundamental determinant of the successful adoption of new educational technologies, which aligns with the findings of Smith et al.¹² and Cabrera et al.¹⁴ on the key drivers of GAI tool adoption in higher education. Notably, a high proportion of users expressed a desire to continue using the platform (Task 7), indicating its perceived sustainability as a long-term learning tool. This convergence of positive subjective experience with objective performance metrics (cognitive load reduction, efficiency improvement, and ESD competency gains) strengthens the internal and external validity of the platform’s design goals and core research findings.

Limitations and Alternative Methodological Approaches
Several limitations of this study must be acknowledged. The participant pool was limited to 50 students and 20 educators from a single institution, which may affect the generalizability of the findings to broader, more diverse populations—a common constraint in initial platform evaluations²³. The study duration, while sufficient for initial adoption, was inadequate to assess long-term educational impact or skill retention. Furthermore, the evaluation primarily relied on quantitative platform metrics and self-reported surveys; deeper qualitative methods, such as longitudinal case studies or design-based research involving iterative co-design with educators^28,29, could yield richer insights into pedagogical transformation and user experience. An alternative approach to testing the core hypothesis could involve a comparative study of different GAI architectural patterns to isolate the most effective components for specific learning outcomes.

Importance, Applications, and Future Directions
The platform’s methodology holds significant importance for specific research and practice areas, directly building on the core findings of this study. In STEM and engineering education, its simulation and immediate feedback capabilities can support complex sustainable design problem-solving, addressing a critical unmet need identified in prior engineering ESD education research²⁷. For professional development, the ESD-aligned competency assessment module enables precise skill-gap analysis for sustainability literacy, a core priority for contemporary engineering training that is rarely integrated into existing GAI educational platforms. The architecture also serves as a vital testbed for learning analytics research, generating rich multimodal datasets to model student behavior, cognitive load, and learning progression, extending established analytical frameworks in the field^18,26. Future work must proceed along four parallel tracks, directly addressing this study’s limitations and building on its core findings: (1) Technical expansion through multimodal AI integration for immersive sustainable engineering design learning, advancing the platform’s capability to support hands-on ESD training aligned with existing technical frameworks¹³; (2) Pedagogical research to define optimal human-AI collaboration models, informed by the varying teacher-AI response ratios observed in our Q&A module, extending related research on educational human-AI interaction^12,15; (3) Large-scale multi-institutional validation across diverse engineering education contexts, to test the generalizability of the platform’s impacts on cognitive load, learning efficiency, and ESD outcomes; and (4) Ethical framework development to address critical issues of algorithmic bias, data privacy, and equitable access, building on established ethical guidelines for AI in education^2,16,24. By pursuing these directions, the potential of GAI to create responsive, effective, and equitable lifelong learning ecosystems that integrate sustainability competency development can be responsibly realized.

Ultimately, this research addresses a critical unresolved gap in GAI educational platform design identified in the extant literature: the lack of a unified, scalable, and pedagogically robust integrated architecture that simultaneously reduces learner cognitive load, improves learning efficiency, and advances ESD outcomes in engineering education. The findings provide rigorous RCT empirical evidence for the transformative potential of intentionally designed, ESD-aligned GAI platforms, and offer a replicable, fully documented architectural blueprint for future research and development in AI-enabled sustainable engineering education.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
A100 GPU (40 GB VRAM)	NVIDIA Corporation	900-4H100-0000-000	Used for hosting core AI models within the server-based processing pipeline.
en_core_web_md (spaCy model)	Explosion AI	https://spacy.io/models/en#en_core_web_md	Pre-trained general academic English model used for tokenization and lemmatization. Version: v3.5.
Python	Python Software Foundation	https://www.python.org/	Programming language for the software environment. Version: 3.9.
SlimPajama dataset	Cerebras Systems Inc.	https://www.cerebras.net/blog/cerebras-releases-slimpajama-a-cleaned-version-of-redpajama/	Large, open-source pre-training corpus derived from RedPajama.
spaCy	Explosion AI	https://spacy.io/	Natural Language Processing library for text preprocessing. Version: 3.5.
TensorFlow	Google LLC	https://www.tensorflow.org/	Open-source machine learning framework. Version: 2.12.