Reinventing Low-Code: Value-Driven and Learning-Oriented

Low-Code Development with SLLM-Integrated Approach

Gayane Sedrakyan

, Stephan Braams

1,2

, Cosmin Ghiauru

, Anton Tsankov

, Stijn Schuurman

Matthijs Jansen op de Haar

, Valeri Andreev

and Jos van Hillegersberg

Department High-Tech Business and Entrepreneurship (HBE) Section, Industrial Engineering and

Business Information Systems (IEBIS), University of Twente, Enschede, Overijssel, Netherlands

Cape Groep, Enschede, Overijssel, Netherlands

Keywords: Low-Code, Model-Based System Development, Business Models, Business Model Canvas,

AI-Enabled Modeling, Requirements Engineering for Low-Code, Citizen Development.

Abstract: Low-code development platforms (LCDPs) are transforming business practices by shifting the focus from

traditional, code-intensive approaches to business-centered modeling. These platforms enable citizen

developers - non-technical employees within organizations - to build and manage applications that address

specific business needs. This democratization accelerates time-to-market and encourages agile, co-

participatory development. However, the rise of citizen development also introduces challenges, such as risks

to quality, security, and governance, due to limited technical expertise among some users. This paper

investigates ways to enhance current low-code practices by integrating AI-based support for text-to-model

generation and established business frameworks, such as the Business Model Canvas (BMC). Incorporating

BMC into low-code platforms reinforces their core strengths - minimizing code dependency while grounding

development in business models. This integration can offer a structured pathway for citizen developers to

engage in meaningful learning while ensuring their projects align with organizational objectives. This

approach positions low-code not only as a productivity tool aiming faster time to market, but as platforms for

continuous learning and strategic alignment with business. The proposed integrations build on a novel

feedback-inclusive approach, which received the innovative feedback nomination at the University of Leuven,

Belgium

, and was informed by evidence-based learning experiences at the University of Twente, Netherlands.

1 INTRODUCTION

The rapid evolution of digital transformation in

business has paved the way for innovative approaches

in software development, among which low-code

development platforms (LCDPs) have emerged as

transformative tools. By reducing reliance on

traditional, code-intensive development, LCDPs

enable a shift toward business-centered modeling,

making application creation more accessible. These

platforms enable re-profiling business developers

into application developers and empower "citizen

developers" - non-technical employees – to actively

participate directly in the process of building and

managing software solutions and prototypes that meet

https://www.kuleuven.be/onderwijs/prijs-onderwijsraad/

2014-2015/Process-oriented-feedback

specific business needs. This democratization of

development not only accelerates time-to-market but

also fosters collaborative, agile processes that engage

a broader range of stakeholders.

Low-code development is a model-based system

development approach that emphasizes creating

applications through enhanced business modeling

and reduced reliance on extensive hand-coding.

In its current form, these platforms offer limited

capabilities for understanding business contexts,

focusing primarily on modeling from requirements

and generating prototypes that can assist in improving

domain knowledge, modeling and modeling language

knowledge that are involved in the development

process with a specific platform. Nevertheless, a

420

Sedrakyan, G., Braams, S., Ghiauru, C., Tsankov, A., Schuurman, S., op de Haar, M. J., Andreev, V. and van Hillegersberg, J.

Reinventing Low-Code: Value-Driven and Learning-Oriented Low-Code Development with SLLM-Integrated Approach.

DOI: 10.5220/0013348200003896

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 13th International Conference on Model-Based Software and Systems Engineering (MODELSWARD 2025), pages 420-431

ISBN: 978-989-758-729-0; ISSN: 2184-4348

substantial amount of tacit knowledge often remains

uncaptured during the transformation of requirements

into models. Enhancing LCDPs with capabilities such

as Text-To-Model Assistance and integration of

Business Model Canvas (BMC) into the LC modeling

phase offers a structured framework to effectively

capture and integrate tacit business insights. This

ensures a more comprehensive representation of

business needs while fostering stronger alignment

between development outputs and strategic business

objectives, ultimately delivering unique / refined

business values. Additionally, this approach aligns

with the model-based development nature of the low-

code development approach. The integration of AI-

based LLMs further advances these capabilities by

enabling more dynamic and context-aware modeling

processes.

In this work, we examine the capabilities and

advantages of incorporating the BMC into the LCD

framework and analyze their impact through

comparative studies. Additionally, integrating overall

modeling support such as text-to-model assistance is

analyzed within the LCD context.

By leveraging the synergy between data, behavior

models and BMC, this research redefines LCDPs as

more than mere productivity tools for rapid

development. It positions them as platforms for

continuous learning and strategic alignment,

enabling organisations to derive sustained business

value and expand development capacity through

citizen-led initiatives. This integration creates a

structured, value-driven and learning-oriented

pathway that enhances the transformative potential of

LCDPs by leveraging their core strengths - business-

centric modeling and the involvement non-technical

developers, such as citizen developers. Additionally,

this approach enriches the learning context for novice

and junior (business) developers, fostering a deeper

understanding of the principles underlying low-code

development and its capacity to generate business

value, while simultaneously enabling practical

experience and skill development through LCDPs.

This research aims to answer the questions:

 RQ1: If and how an AI-based text-to-model

assistance can contribute to the knowledge

enhancement for LCD?

 RQ2: If and how the integration of Business

Model Canvas into the Low-Code modeling

cycle enhances knowledge on capturing added

business value ?

This paper is organized as follows: Chapter 2

provides an overview of LCDPs, including their

theoretical background, typologies, commonly used

models, and an analysis of their benefits and

limitations. Chapter 3 examines the integration of

learning support within the low-code development

lifecycle, incorporating a text-to-model approach and

the concept of business value into the modeling cycle

through generative AI. Chapter 4 presents the results

of case studies comparing the quality of low-code

models and applications developed with unassisted

and AI-assisted cycles. It also discusses the findings

and highlights the limitations of the study. Finally,

Chapter 5 concludes by evaluating the impact of the

integration of AI-based support described in the study

on enhancing low-code models to support for

improved knowledge of LCD developers with

particular focus on novice and citizen developers,

along with recommendations for future research.

2 RELATED WORK

Low-code development platforms enable users to

create applications with minimal hand-coding

(Sedrakyan & Snoeck, 2013), relying on visual

modeling, drag-and-drop interfaces, and pre-built

components. This aligns with the principles of

Model-Based Systems Engineering (MBSE), where

models represent system architecture and behavior

(Di Ruscio et. al, 2022). In the context of low-code,

the visual models serve as both the blueprint and

executable logic, offering a unified framework for

system design. This chapter is structured as follows:

(1) a summary of the theoretical foundations of low-

code development, (2) an exploration of low-code

model types and notations commonly used in

development processes, and (3) support for learning

context, and (4) a description and a discussion of the

Business Model Canvas and its capabilities.

2.1 LCD Definitions and Typology

LCD is characterized by its use of visual, declarative

techniques and minimal hand-coding, offering two

primary approaches:

 Descriptive low-code;

 Prescriptive (composable) low-code;

 No-Code development;

 AI-Enhanced Low-Code;

 Hybrid approaches.

The descriptive low-code approach allows

developers to visually design the structure and

functionality of applications, often leveraging

graphical interfaces and drag-and-drop components

to represent software elements. Mendix is an example

of such descriptive LC platform (Henkel & Stirna,

2010). In contrast, the prescriptive (or composable)

Reinventing Low-Code: Value-Driven and Learning-Oriented Low-Code Development with SLLM-Integrated Approach

421

low-code approach focuses on rapidly assembling

applications using pre-built components and

templates, reducing the need for custom coding.

Novulo is an example of a composable LCDP

. No-

code development approach uses a subset of LCD

designed exclusively for non-technical users,

enabling application development without any

coding. These platforms rely entirely on pre-defined

templates, visual interfaces, drag-and-drop and pre-

built components with automated workflows, making

them ideal for citizen development initiatives. AI-

enabled LC combines the principles of low-code with

AI-based tools, such as large language models

(LLMs), to support more dynamic development

processes. These platforms can assist in generating

models, code, or application logic based on natural

language descriptions or domain-specific inputs,

significantly reducing the cognitive load on

developers. Hybrid approaches combine the

strengths of low-code platforms with traditional high-

code (HC) approaches, enabling developers to

leverage the best potential of both domains. This

hybrid approach is becoming increasingly popular as

it addresses the limitations of pure low-code or high-

code systems, allowing benefiting from the speed and

simplicity of low-code platforms while retaining the

flexibility and precision of high-code development.

There are two main approaches for LC and HC

integrations:

 Extending Low-Code with High-Code: Low-

code platforms are often limited in their ability

to handle highly customized features or

integrations with complex systems. By

integrating high-code, developers can extend

the functionality of low-code applications, such

as implementing custom algorithms,

integrating advanced APIs, or designing unique

user interfaces that go beyond the capabilities

of visual tools. As an example, a low-code

CRM system could be extended with high-code

to include a custom AI-based recommendation

engine for personalized sales strategies.

 Using Low-Code in High-Code Applications:

Conversely, low-code components can be

embedded within high-code applications to

accelerate development cycles for less critical

or repetitive modules. For instance, low-code

tools can be used to quickly prototype

dashboards, automate workflows, or build

internal tools, reducing the burden on high-

code developers. As an example, a high-code e-

commerce platform could incorporate a low-

https://www.novulo.com/

code module to manage and automate

inventory processes, allowing business

analysts to make changes without requiring in-

depth programming expertise.

The user base for low-code platforms is diverse,

including business analysts / developers,

experienced software engineers, citizen

developers. Among typical business use cases for

low-code applications are:

 Rapid Prototyping: Quickly testing ideas and

concepts to validate them before significant

investment.

 Legacy System Modernization: Addressing

outdated systems with scalable, flexible

solutions.

 Process Automation: Streamlining repetitive

tasks and workflows.

 Citizen Development Initiatives: Empowering

non-technical users to participate in and/or

build applications tailored to departmental

needs, reducing IT workloads.

2.2 Theoretical Background of LCD

The theoretical foundation of low-code traditionally

includes model-driven engineering (MDE),

providing a framework for generating software

applications based on highly abstracted models and

(meta)model-to-(meta)model transformations

(Sedrakyan & Snoeck, 2014). MDE principles guide

the creation of reusable models and transformations

that capture domain-specific knowledge and facilitate

code generation with minimal manual effort.

Descriptive low code leverages MDE to generate

software applications based on modeled data,

behavior, and user interfaces visually. Prescriptive

low code, focused on modular design and component

reuse, aligns with MDE principles through domain-

specific modeling languages (DSMLs) for modeling

LC applications. Rather than relying on structural and

behavioral models of a system, this approach focuses

on integrating reusable components through models

of reusable components that allow "stitching" these

components together to enable the rapid assembly of

applications with enhanced consistency, reusability,

and scalability.

2.3 Common Model Types and

Notations Used in LCD

The development lifecycle of the current descriptive

low-code (DLC) approach, which will be the focus of

MBSE-AI Integration 2025 - 2nd Workshop on Model-based System Engineering and Artiﬁcial Intelligence

422

this research, is depicted in Figure 1. It typically

begins with capturing the data requirements of the

information system, followed by designing

workflows and defining detailed functionalities.

Figure 1: Common Modeling Workflow in Descriptive

Low-Code Development Platforms.

Next, the user interface design and interactions are

established, culminating in the generation of a fully

functional application.

LCDPs can employ a variety of model types and

notations to facilitate visual application design.

These include models for structured

representations of application components,

behaviors, and interactions:

 Structural Models define the underlying data

structures and relationships within the

application. Commonly used notations include:

o UML Class Diagrams to represent

Domain Models through data structure

representing entities, attributes, and their

relationships (e.g. database schema).

o Entity-Relationship Diagrams depicting

data relationships in database-centric

designs.

 Behavioral Models focus on the dynamic

aspects of the application, such as workflows

and event-driven processes. Notations often

include:

o BPMN (Business Process Model and

Notation) visualizing business

workflows and decision points.

o Activity Diagrams allowing visualize

tasks, conditions, and parallel processes

to define user interactions and business

logic visually, e.g. handling data

validation or conditional routing.

o UML State Diagrams (or Statecharts)

capturing state transitions based on user

actions or system events, e.g. handling

state changes for a ticket system (e.g.,

"Open", "In Progress", "Closed").

o Use Case Diagrams defining user roles

and interactions to support user stories,

e.g. mapping out user roles like "Admin"

and "Customer" and their interactions

with the app.

 User Interface Models describe the design and

interaction patterns of the application’s front-

end, often including:

o UI Flow Diagrams illustrating user

navigation paths with a high-level

overview of how users navigate through

the application, including pathways

between pages, screens, or modal

windows.

o UML State Diagrams depict how the

application responds to user input by

transitioning between different states.

For instance, a login form might

transition to a "logged in" state upon

successful authentication.

o Wireframes and Mockups allowing

sketching the layout and navigation.

o Drag-and-Drop Interfaces are widely

adopted in LCDPs, often with "what-

you-see-is-what-you-get" (WYSIWYG)

editors that allow LC developers to

design interfaces visually, reducing the

need for manual coding expertise.

Hybrid approaches combine models are

often used in combination to map out the

flow of interactions between the user and

the application, aiming to capture how

different elements of the interface

respond to user actions.

 Component Models are used in prescriptive

low-code approaches to define reusable

building blocks using module libraries

encapsuling pre-built functionalities and

services.

2.4 Transformative Benefits and

Drawbacks of LCD

LCDPs represent a paradigm shift in software

development, offering significant opportunities for

organizations to enhance efficiency, reduce costs, and

increase adaptability. By abstracting the complexity

of traditional development processes through models,

LCDPs enable broader participation in software

creation and facilitate rapid prototyping and

deployment. However, their adoption introduces

certain technical and organizational challenges that

necessitate critical evaluation (Rokis & Kirikova,

2022). This section outlines the benefits and

challenges of LCD, emphasizing their implications

for business value and operational dynamics.

LCD offers transformative benefits for software

development by enabling faster, more adaptive

processes that align with evolving business needs.

Reinventing Low-Code: Value-Driven and Learning-Oriented Low-Code Development with SLLM-Integrated Approach

423

Although not exhaustive, the following benefits

highlight key advantages:

 Cost Efficiency and Democratized

Development: LC approach empower non-

technical personnel, or citizen developers, to

participate in application development,

reducing reliance on specialized expertise. This

democratization not only lowers development

costs but also expands organizational capacity

for addressing diverse business needs through

software solutions.

 Accelerated Time-to-Market: LC approach

leverages visual modeling and reusable

components to significantly reduce

development timelines. This capability

supports rapid prototyping and the iterative

refinement of Minimum Viable Products

(MVPs), fostering innovation within tight

timeframes.

 Testing Ideas with Uncertain Requirements:

The visual design tools and rapid development

cycles of LCDPs facilitate continuous feedback

loops, enabling early and frequent

incorporation of stakeholder insights. In

dynamic environments, where requirements

are frequently incomplete or subject to change,

LCDPs provide the flexibility needed for

iterative development. This adaptability

reduces the risks of developing misaligned or

obsolete solutions. With this capability LCDPs

align with the "Mode 2" of the frameowork

called “BimodalIT”

, which emphasizes

experimentation and responsiveness in

contexts where requirements remain fluid

(Horlach et al., 2022). By providing flexibility

for iterative development, LCDPs are well-

suited for dynamic environments where

requirements are incomplete or subject to

change, supporting organizations in testing

ideas and responding to shifting market

demands while maintaining stability balanced

with experimentation and adaptability.

 Agile Development and Enhanced

Collaboration Across Teams and with

End-users / Customers: By offering a faster

development cycles to deliver MVPs, LCD

promotes shared understanding between

technical teams and business stakeholders. This

collaboration bridges the traditional divide

between IT and business functions, enabling

more precise translation of strategic objectives

https://www.gartner.com/en/information-technology/

glossary/bimodal

into functional software, while also integrating

end-users into iterative prototyping cycles.

This collaborative and agile approach not only

improves development efficiency but also

enhances customer satisfaction through their

involvement in earlier phases of development,

continuous feedback and refinement.

 Legacy System Modernization and fast

pluggable services: Organizations often face

challenges in updating legacy systems that are

integral to their operations. LCDPs provide

scalable solutions that extend the functionality

of outdated systems, integrating them with

modern software infrastructures and reducing

the need for costly overhauls.

 Maintainability: LCDPs simplify updates

through visual modeling and reusable

components, facilitating rapid adaptation to

changing requirements with centralized

management. For example, modifying a model

and re-deploying is typically sufficient to

introduce changes to the LC applications.

Integrated Continousu Integration and Delivery

(CI/CD) pipelines (Humble & Farley, 2010)

optimize deployment processes through

reduced downtime and enhanced operational

agility.

 Cross-platform Deployment: LCDPs support

responsive designs and native mobile

capabilities, allowing applications to function

across devices and platforms.

While not exhaustive, the following challenges

highlight key drawbacks with LCD approach:

 Potential for Errors by non-IT developers:

Citizen developers may lack technical skills,

the domain-specific and modeling and

modeling language expertise to fully

understand or translate complex requirements

into accurate models, leading to inconsistencies

and errors.

 Security Vulnerabilities: The abstraction

inherent to LCDPs and lack of technical insight

of non-IT developers can obscure critical

security considerations, increasing the risk of

vulnerabilities in the resulting applications

(Sedrakyan at al., 2024).

 Scalability and Customization Constraints:

While LCDPs excel in rapid and standardized

application development, their capacity to

address highly customized or large-scale

enterprise requirements is often limited,

MBSE-AI Integration 2025 - 2nd Workshop on Model-based System Engineering and Artiﬁcial Intelligence

424

necessitating integration with high-code

solutions, increasing complexity.

 Dependence on Platform Ecosystems: The.

reliance on proprietary tools and environments

introduces significant risks, including vendor

lock-in, platform discontinuity, or deprecation

of features. These risks can restrict long-term

strategic flexibility and make migration to

alternative solutions costly and challenging.

 Reliance on Third-party Tools and Cloud

Servers: Many LCDPs depend on third-party

integrations, which may use outdated or

unsupported versions of libraries, APIs, or

plugins. This dependency can introduce

vulnerabilities, compatibility issues, or

maintenance challenges over time.

Additionally, reliance on cloud-based servers

can create privacy and security concerns,

especially when sensitive data is stored or

processed in environments that may not fully

comply with regulatory or organizational

standards. This can increase the risk of data

breaches, unauthorized access, or compliance

violations.

3 LEARNING- AND

VALUE-ORIENTED

ENHANCEMENTS

This chapter delves into the enhancement of the LCD

lifecycle with a learning support. It investigates if and

how the integration of AI-driven advancements can

influence knowledge gaps. It explores how feedback-

enabled LCD cycles can refine iterative design

processes, ensuring adaptability and continuous

improvement. Additionally, the chapter examines the

possibilities for transformation of textual inputs into

structured data and process models and the

incorporation of business value frameworks into the

modeling process, demonstrating the potential for AI

to bridge knowledge gaps of novices in the context of

LC application development and output quality.

3.1 Model-Based Feedback Integration

As models serve as the central engine of LCDPs, the

quality of these models is critical for generating

effective applications. Modeling, however, is a

multifaceted skill that extends beyond theoretical

understanding, requiring practical expertise. While

LCDPs simplify the learning curve through visual and

intuitive interfaces, making them better suited for

business and citizen developer profiles, learning

processes can be significantly enhanced by feedback-

enabled mechanisms. Such mechanisms not only

improve the transformation of requirements into

functional applications but also foster iterative

refinement, enabling trial-and-error cycles that

promote deeper learning and improved model

accuracy. Since modeling forms the foundation of

low-code approaches, embedding learning aids early

in the modeling process ensures users develop a

stronger grasp of the concepts and techniques.

Existing research underscores the value of model-

based automated feedback in advancing novice

modelers' skills across various stages of system

design. Feedback mechanisms linking the execution

outcomes in a resulted application to the parts of

models that cause them have proven effective in

refining structural models (Sedrakyan & Snoeck,

2013), e.g., understanding and improving the semantic

quality of UML class diagrams (Sedrakyan & Snoeck,

2017), improving behavioral models, e.g., UML

statecharts (Sedrakyan et al., 2017), to address event

failures, and iteratively guiding user interface design

(Ruiz et al., 2015). Furthermore, feedback-driven

testing assists in defect detection and validates

requirements against model specifications, mitigating

risks associated with limited domain or technical

knowledge (Sedrakyan & Snoeck, 2014).

Additionally, modeling behavior patterns were found

to be linked with modeling process outcomes,

suggesting that incorporating feedforward

mechanisms to stimulate effective modeling behaviors

has the potential to enhance model quality (Sedrakyan

& Snoeck, 2017).

Given the model-centric nature of LCDPs,

integrating model-driven feedback mechanisms aligns

well with their core objectives and design principles.

Such integration has the potential to enhance learning

outcomes and development efficiency, expanding the

scope of LCDPs to leverage citizen development more

effectively in areas previously considered better suited

for IT expertise to better respond to the IT talent

deficit. Research in educational contexts further

supports the effectiveness of model-driven feedback

for learners with profiles similar to citizen developers,

emphasizing its potential to empower these users and

improve their ability to contribute meaningfully to

development processes.

3.2 LLM-Based Text-to-Model (T2M)

Assistance

In this study, we employ an integrated descriptive

Low-Code (LC) and AI-based approach. Descriptive

Reinventing Low-Code: Value-Driven and Learning-Oriented Low-Code Development with SLLM-Integrated Approach

425

LC was selected due to its effective alignment with

Information Systems data structures and business

processes within a modeling context. The Text-to-

Model (T2M) approach was implemented using the

TeToMo framework (Sedrakyan et al., 2022).

Integration was achieved through a ChatGPT

(OpenAI, 2024) API connector for Mendix

. The

query used to interact with ChatGPT is as follows:

'Generate MermaidJS code for a '

+replaceAll(toLowerCase(toString($Ch

atGPTRequest/Dropdown)),'_','') + '

+ $ChatGPTRequest/Prompt

This query utilizes the selected diagram type from a

dropdown menu and constructs a prompt for

generating appropriate MermaidJS code. The output

diagram is derived from the input text specified in

$ChatGPTRequest/Prompt. Java code is then used

to format the ChatGPT response into usable

MermaidJS code. Afterwards a microflow designed

in Mendix Modeler, is called to processes user input

and format it. Subsequently, a second microflow is

invoked to utilize the ChatGPT API and retrieve a

response. The resulting response, in JSON format, is

processed by Java code to generate the final diagram.

This diagram is stored as a new "Diagram" entity and

displayed on the user page.

3.3 Integrating Business Model Canvas

Using an SLLM Approach

While the quality of models is central for the quality

of LCD applications, in the modern competitive

landscape, success is not solely measured by

technical accuracy between requirements and

application design artefacts (models). The ability to

create and capture business value has become a

critical differentiator. Enhancing LC platforms with

feedback systems that can assist with capturing and

modeling business value can bridge this gap,

empowering citizen developers and novice users to

align application development with strategic business

objectives. This dual focus on learning and value

creation has the potential to position LC platforms as

not only productivity tools but also as enablers of

innovation and competitive advantage. Enhancing LC

models with capabilities for capturing business value

in addition can further amplify its learning context,

offering developers the tools to bridge technical and

business perspectives. This integration could

significantly expand the LC platform’s role not only

https://marketplace.mendix.com/link/component/206616

as a productivity but also as an educational

environment.

The Business Model Canvas (BMC) is a strategic

management framework used to visualize, design,

and refine business models to capture business value

and align operations around value creation. It consists

of nine interconnected building blocks: Customer

Segments, Value Propositions, Channels, Customer

Relationships, Revenue Streams, Key Resources,

Key Activities, Key Partnerships, and Cost Structure.

These elements collectively represent the logic of

how an organization creates, delivers, and captures

value. As LCD platforms focus on enhancing

business models and reduce manual coding for

accelerated application development by empowering

citizen developers, integrating business-focused

frameworks is crucial to align with this goals. The

Business Model Canvas (BMC), with its emphasis on

value creation and strategic alignment between

organisational objectives and market demands, is

particularly well-suited to this integration.

This study introduces the term Specific Large

Language Model (SLLM) to describe LLMs tailored

for specific domains or tasks. The integrated

descriptive modeling with AI-assisted BMC is

achieved through a Specific Language Model

(SLLM) chatbot designed for integration with the

Canvas Learning Management System (CLMS)

the University of Twente (UTwente). The chatbot

aims to improve accuracy by also learning from

CLMS courses and offer specialised support to

students for their tasks that link to teacher-specified

resources. The integration is using RAG (Retrieval-

Augmented Generation), a technique that combines

retrieval of relevant documents from a knowledge

base with the generation of responses, enhancing the

quality of output in conversational AI systems (Cai et

al., 2022).

To ensure that sensitive information, such as

student and teacher data, is not transmitted to external

AI providers like OpenAI or Meta via internet APIs,

the SLLM is hosted in a local Docker container. The

system uses the Llama3.1-8b-instruct-q8 (Ollama,

n.d.) which is a 8-bit quantized Llama. While more

powerful model versions are available, they demand

substantially more system resources to improve

response precision and accuracy. We have chosen the

8-bit quantized model for its resource efficiency and

adequate performance. Python server in FastAPI

(Rodríguez, n.d.) is used for routing the requests of

the conversational chatbot to the appropriate services.

https://www.utwente.nl/en/educational-systems/about-the

-applications/canvas/

MBSE-AI Integration 2025 - 2nd Workshop on Model-based System Engineering and Artiﬁcial Intelligence

426

the /load request that utilizes the UTwente’s Canvas

API to gather the required modules, documents and

pages from the given course, converts it to text,

chunks and embeds the pieces of texts. These

embedded chunks are then stored inside the Weaviate

vector database

. The /status request is then used to

verify whether a database entry exists for a given

course, ensuring compatibility of the extension for

student use. Subsequently, the /chat_stream request

is employed to handle questions posed by students

within a course. This request processes the user’s

query and its history, reformulating the query into a

contextual format to perform a similarity search

within the Weaviate database. After receiving the

response from Weaviate, the query, along with the

retrieved documents, is forwarded to the LLM for

further processing and response generation. Many

course files on Canvas are in PDF format, and text

extraction is natively supported by the Python code

on the server. However, for non-native formats such

as Microsoft Office files (e.g., Word and

PowerPoint), LibreOffice

, an open-source office

software, is employed to ensure compatibility.

Teachers can utilize domain-specific prompts to

narrow down the guidance provided by the system,

ensuring alignment with course objectives and

desired learning outcomes. The resulting SLLM

enables students to receive answers not only sourced

from the internet but also refined with tailored

knowledge, course-specific resources, and contextual

information derived from historical learner data, thus

supporting more accurate and personalized learning

assistance.

3.4 Learning Effects of T2M and BMC

Interventions

This section explores the outcomes of two quasi-

experimental studies designed to evaluate the impact

of different intervention methods on novice learners'

ability to develop Low-Code (LC) applications.

Unlike traditional experimental designs that typically

involve the controlled manipulation of multiple

variables to mitigate the influence of unknown

factors, the quasi-experimental approach adopted

here reflected a more natural and flexible design. This

approach aligns with real-world scenarios where

controlling all variables is challenging, as in our case,

requiring incremental development cycles spread

over the course timeframe. Yet, without strict

manipulation of multiple variables, the findings

provide corrections for which the effects of unknown

https://weaviate.io/

variables, although important in traditional

experimental research, were considered less critical to

the outcomes of these studies.

It is important to note that the tools described in

Sections 3.2 and 3.3 were not utilized during these

quasi-experiments, as they were still under

development. Instead, basic AI-based assistance

using ChatGPT was employed. The subsequent tool

development described earlier was informed by the

findings of these quasi-experiments, aligning with the

interventions applied and aimed at advancing toward

a specialized, integrated, all-in-one environment.

This development also incorporated evidence-based

perspectives from learning sciences, such as feedback

typology and formats, as well as adherence to

technological standards, including the safeguarding

of learner data privacy.

The quasi-experiments employed randomized

student samples and course project topics, examining

interventions across two cycles: an unassisted cycle

(referred to as the 1st cycle) and an assisted cycle

using ChatGPT-based tools (referred to as the 2nd

cycle). In the first quasi-experiment (referred to as the

1st experiment), text-to-model assistance was

introduced during the 2nd cycle to support the

modeling of system structure and behavior within the

Mendix modeler. The second quasi-experiment

(referred to as the 2nd experiment) implemented AI-

based BMC feedback during the 2nd cycle to enhance

application refinement.

The learning context and task descriptions

included five cases of similar complexity, which were

offered to students to select from and develop into

applications as part of their course project tasks.

These cases included requirements descriptions that

necessitated further elicitation from end-users (e.g.,

company use case owners or teachers acting as case

owners). Examples of the case descriptions included

sustainable food recommender applications, fitness

recommender apps, and learning support

applications.

A total of 22 student outcomes were examined

during the academic year 2022–23 (1st experiment),

and 19 student outcomes during the academic year

2023–24 (2nd experiment). The participants

represented diverse backgrounds, spanning both

technical and non-technical master-level study

programs, and ranged in age from 22 to 35.

Performance was evaluated using cumulative rubric

points for intermediate and final outcomes.

https://www.libreoffice.org

Reinventing Low-Code: Value-Driven and Learning-Oriented Low-Code Development with SLLM-Integrated Approach

427

Table 1: Corrections observed after the assisted cycle in

Experiment 1 (CX1, Column 2) and Experiment 2 (CX2,

Column 3).

MVP

ualit

assessment metrics* CX1 CX2

ture o

business context 1 3

Customer Journey and Business

Processes

3 4

dded value

or business 1 3

ddresses customer

roblem 1 4

Link between business processes and

value

2 5

Data model

ualit

and desi

n choices 3 1

ctors & Roles 2 3

Security risks identified and addressed 1 0

Overall MVP quality (all features

included are essential

or the business

)

2 5

Validation with users

(

end-user score

)

1 3

UI (intuitive navigation) 0 0

Behavior validation for desired and

undesired

aths

1 0

* All criteria scores were rounded to the nearest integer

In the 1st experiment, students were required to

transfer their business requirements into LC

applications during an unassisted cycle (1st cycle).

They subsequently utilized a T2M assistance cycle

with a ChatGPT version to enhance their models (2nd

cycle). In the 2nd experiment, students developed

their chosen cases into LC applications during the

unassisted cycle and later refined their applications

using feedback from intermediate results in a BMC-

integrated cycle.

The intermediate (unassisted) and final (assisted)

quality of the developed applications was evaluated

using an assessment rubric covering multiple criteria

(Table 1, column 1): capturing business context;

customer journey and business processes; added

value for business; addressing customer pain points;

linking business processes to value; data model

quality and design choices; actors and roles;

identification and mitigation of security risks; overall

Minimally Viable Product (MVP) quality (ensuring

all features were essential for the business); validation

with end-users (denoting customer feedback scores);

UI intuitiveness for navigation; and validation of both

desired and undesired paths. Application quality was

rated by experts on a 0–10 scale, with 0 indicating an

unsatisfactory solution and 10 representing an

excellent solution meeting all rubric criteria.

In the 1st experiment, as model quality was a

major objective, specific sub-criteria were used to

evaluate the data model and design choices. These

included the number of correct/incorrect entities,

attributes, associations, actors and roles, and

activities; semantic quality in terms of accurately

reflecting the requirements; and valid

desired/undesired paths (e.g., modeled constraints),

application of best practice patterns. Both positive

and negative corrections were analyzed after the

intervention. The 1st experiment demonstrated a

positive correction of 1.5 after employing AI-based

T2M assistance (see the average for each graded

criterion in Table 1). No substantial negative

corrections were observed in any of the rubric criteria

or sub-criteria related to model quality. No substantial

negative corrections were observed in any of the

rubric criteria or sub-criteria related to model quality

for the 2

experiment either.

Normality testing revealed both normal and non-

normal distributions in both experiments, which were

appropriately assessed before and after the

interventions. These findings aligned with the

anticipated corrections resulting from the applied

interventions.

4 DISCUSSION

This section synthesizes the findings from the two

quasi-experimental studies conducted to evaluate the

effects of different intervention methods on novice

learners' abilities to develop LC applications. The

study aimed to assess the impact of AI-based

interventions, specifically Text-to-Model assistance

using ChatGPT and AI-based BMC creation

assistance - on the quality of resulting LC

applications. The findings from both experiments

underscore the effectiveness of AI-assisted tools in

improving the quality of LC applications developed

by novices. The positive corrections observed in both

experiments align with existing research suggesting

that AI-based assistance can provide valuable

guidance within LCDPs as learning environments,

particularly for tasks that require complex problem-

solving such as requirements elicitation and

engineering requiring modeling skills. The results are

consistent with studies that highlight the potential of

assisted modeling environments to improve learning

outcomes by offering personalized and context-aware

feedback (Sedrakyan & Snoeck, 2016). The lack of

negative corrections further suggests that the

intervention was beneficial, without introducing any

unintended consequences.

The T2M approach exhibited significant potential

in assisting students to identify key model elements,

such as entities, attributes, relationships, mandatory

and optional associations, events, and their

sequences, derived from textual requirements. The

second experiment, which incorporated AI-based

BMC assistance, further demonstrated the

MBSE-AI Integration 2025 - 2nd Workshop on Model-based System Engineering and Artiﬁcial Intelligence

428

effectiveness of AI in supporting novice developers

in business-oriented tasks, bridging the gap between

requirements modeling and business strategy to

create more meaningful (LC) applications.

Applications and learning cycles for novice

developers were positively evaluated by domain

experts, including educators, LC platform vendors,

and consultants with extensive experience in LC

development. These findings underscore the potential

of integrating AI-driven Text-to-Model and Business

Model Canvas integration assistance to enhance the

functionality and educational value of LC platforms.

This enhancement aligns with the fundamental

characteristics of LC focusing on increasing the

emphasis on business modeling while reducing

reliance on manual coding efforts. Furthermore, it

reinforces the objectives of LC development to

enhance on citizen development and business value

generation.

Interestingly, slightly higher corrections were

observed among students enrolled in programs

lacking prior knowledge of UML and programming,

suggesting that the approach effectively supports the

knowledge-building processes of novices. However,

students with technical backgrounds consistently

outperformed their non-technical peers even in the

AI-assisted cycle. This observation suggests that,

while the LCD approach minimizes the required

technical expertise, developers with technical insight

still derive additional benefits.

While the quasi-experimental design provided

valuable insights, several limitations should be

considered. The study was conducted with a sample

of students from various academic backgrounds and

years, which, while contributing to the

generalizability of the findings, may have influenced

the results. Additionally, it is difficult to determine

whether the observed improvements were solely due

to the interventions or if other factors played a role.

Future studies could explore the impact of AI-based

interventions in more homogeneous student groups to

better isolate the effects of the interventions.

Furthermore, experiments involving heterogeneous

samples with stricter designs could help measure the

effects of compound or unknown variables, thus

providing a more comprehensive understanding of

the interventions. Testing the effects with larger

samples across different academic institutions or even

with junior professionals from industry and citizen

developers could further enrich the findings.

Furthermore, while the interventions showed

positive effects on model quality and application

development, it remains unclear whether these

improvements translate into long-term learning gains

or real-world application success. Longitudinal

studies that track the retention of skills and the real-

world performance of the applications developed by

students would be valuable in assessing the lasting

impact of AI-based interventions.

4.1 Considerations

Another important point for discussion is the

accuracy of the AI-based assistance used during the

experiment, particularly regarding hallucination, a

common challenge for generative AI systems. While

current results are promising, future implementations

can enhance reliability by leveraging Retrieval-

Augmented Generation (RAG). RAG reduces

hallucinations by grounding responses in course

content, ensuring outputs are both accurate and

contextually relevant. For instance, prompts within

the T2M and BMC systems can be adapted to

educational contexts by embedding references to

specific course materials or assessment criteria. This

approach not only aligns AI assistance with cognitive

and socio-cognitive learning theories.

Equally significant is the innovative and

sustainable concept of decentralized, which could

transform the feedback architecture. This approach

envisions deploying SLLMs locally on user devices,

leveraging concepts from federated learning. In such

a setup, models can be periodically updated and

trained locally, with aggregated insights shared back

to a central system without transferring raw data. This

not only ensures privacy but also minimizes server

reliance, promoting scalability and sustainability.

Technically, this will involve lightweight,

containerized deployments of SLLMs on user

machines, with mechanisms for secure

synchronization and conflict resolution when

aggregating updates. Key considerations include

optimizing resource use on user devices, ensuring

model coherence across decentralized systems, and

maintaining robust encryption to protect sensitive

data during synchronization.

5 CONCLUSIONS

The findings of this study demonstrate the potential

of AI-based assistance in improving novice learners'

abilities to develop Low-Code applications. The

Text-to-Model assistance and BMC feedback

interventions led to significant improvements in both

the technical quality and business aspects of the

applications, underscoring the value of AI in

supporting a comprehensive perspective of

Reinventing Low-Code: Value-Driven and Learning-Oriented Low-Code Development with SLLM-Integrated Approach

429

developing with Low-Code platforms in a learning

context (RQ1, RQ2). The study found that students

with no prior knowledge of modeling and

programming showed slightly greater improvements

in their assisted cycle, indicating that the approach

effectively supports the knowledge-building

processes of novices (RQ1, RQ2). However, students

with technical backgrounds continued to outperform

their non-technical peers, even in the AI-assisted

cycle (RQ1, RQ2). This suggests that, while the LCD

approach reduces the technical expertise required,

developers with technical insights still perform better

and are likely to derive more benefits from AI due to

their background knowledge (RQ1, RQ2). The

findings of this study show that AI-based text-to-

model support, is helpful in identifying and justifying

potential model elements such as classes, attributes,

associations, and activities as evidenced by

improvements in application quality (RQ1). The

study also highlights the potential of integrating the

Business Model Canvas into the low-code (LC)

development approach, demonstrating its capacity to

enhance the model-based development approach of

LC by incorporating business value modeling, as

evidenced by improvements in application quality

(RQ2). Beyond serving as a productivity tool to

accelerate development speed, reduce time-to-

market, and lower costs, LC platforms augmented

with the BMC also offer a capacity to serve a learning

environment. This dual functionality underscores the

transformative role of LC platforms in fostering both

rapid prototyping and developer skill acquisition

through trial-error loops.

Future research directions include integrating the

suggested AI-based approaches to refine LC

capabilities even further. The findings of the

experiments highlight the need for addressing

security risks, validation mechanisms, and user

interface design in the modeling process. The

RAAFT framework proposed by Sedrakyan et al.

(2024) provides a foundation for designing security

modeling guidelines that can be effectively integrated

into low-code (LC) models. Additionally, enhancing

modeling support for validation, such as prompts for

(assisting) detecting undesired paths with model

generation assistance (e.g. microflow within the

Mendix modeler), offers a promising avenue for

future research.

Furthermore, the integrated BMC and T2M

approach outlined in this study has applications both

in bridging the LCD loop through a generative text-

to-application capability and beyond, demonstrating

potential for supporting assessment assistance in

educational contexts, such as enabling automated

grading systems.

To evaluate the practical impact, other potential

future directions for this study include studies with

broader participant pools, including both academic

and industry settings, with stricter experimental

designs that account for effects from other variables

(such as learning from the case in the first cycle) and

compound unknown variables.

While the quasi-experimental design in this study

did not offer strict control over all contributing

variables, the positive outcomes suggest that AI-

enhanced interventions integrated into low-code

platforms are promising. Future research should

further investigate the educational role of AI, with an

emphasis on broadening the scope of educational

interventions and evaluating long-term learning

outcomes. Given that model quality scores may not

accurately reflect learning processes and skills

acquisition, alternative assessment criteria warrant

exploration. To further enhance LCDPs, integrating

cognitive feedback and behavioral feedforward

mechanisms, as suggested by Sedrakyan et al. (2016),

could provide enriched learning and development

experiences within the LC context. Additionally,

enhancing model-based mechanisms and AI-based

support for other phases of the LC lifecycle, such as

guiding the service and third-party API integrations

with models that enable triggering/calling

integrations, and offering guidance for platform

selection, would enhance the versatility and

functionality of LCDPs. These advancements,

alongside deeper AI integrations, have the potential

to elevate LCD platforms, making them more

powerful and adaptable for a broader range of users

and use cases. While these developments have the

potential to redefine LC platforms, making them

more accessible and versatile for a broader range of

users, future research must carefully address

scenarios where human input and judgment remain

indispensable to ensure reliability, accuracy, and

alignment with complex business objectives.

Additionally, incorporating mechanisms for

preventing deskilling due to excessive reliance on AI-

driven assistance represents an important avenue for

future study.

ACKNOWLEDGEMENTS

The research is conducted in the context of the course

Low Code Application Development: UTwente

OSIRIS site at https://utwente.osiris-student.nl/

onderwijscatalogus/extern/cursus.

MBSE-AI Integration 2025 - 2nd Workshop on Model-based System Engineering and Artiﬁcial Intelligence

430

REFERENCES

Sedrakyan, G., & Snoeck, M. (2013, February). A PIM-to-

Code requirements engineering framework. In

International Conference on Model-Driven

Engineering and Software Development (Vol. 2, pp.

163-169). SciTePress.

Sedrakyan, G., & Snoeck, M. (2014). Lightweight semantic

prototyper for conceptual modeling. In Advances in

Conceptual Modeling: ER 2014 Workshops, ENMO,

MoBiD, MReBA, QMMQ, SeCoGIS, WISM, and ER

Demos, Atlanta, GA, USA, October 27-29, 2014.

Proceedings 33 (pp. 298-302). Springer International

Publishing.

Di Ruscio, D., Kolovos, D., de Lara, J., Pierantonio, A.,

Tisi, M., & Wimmer, M. (2022). Low code

development and model-driven engineering: Two sides

of the same coin?. Software and Systems

Modeling, 21(2), 437-446.

Henkel, M., & Stirna, J. (2010). Pondering on the key

functionality of model driven development tools: The

case of mendix. In Perspectives in Business Informatics

Research: 9th International Conference, BIR 2010,

Rostock Germany, September 29–October 1, 2010.

Proceedings 9 (pp. 146-160). Springer Berlin

Heidelberg.

Horlach, B., Drews, P., & Schirmer, I. (2016, March).

Bimodal IT: Business-IT Alignment in the Age of

Digital Transformation. In MKWI (pp. 1417-1428).

Rokis, K., & Kirikova, M. (2022, September). Challenges

of low-code/no-code software development: A

literature review. In International Conference on

Business Informatics Research (pp. 3-17). Cham:

Springer International Publishing.

Sedrakyan, G., Iacob, M. E. , & van Hillegersberg, J. 2024.

Towards LowDevSecOps Framework for Low-Code

Development: Integrating Process-Oriented

Recommendations for Security Risk Management. In

Proceedings of the ACM/IEEE 27th International

Conference on Model Driven Engineering Languages

and Systems (MODELS Companion '24). Association

for Computing Machinery, New York, NY, USA, 886–

894. https://doi.org/10.1145/3652620.3688335

Humble, J., & Farley, D. (2010). Continuous delivery:

reliable software releases through build, test, and

deployment automation. Pearson Education.

Sedrakyan, G., & Snoeck, M. (2013, June). Feedback-

enabled MDA-prototyping effects on modeling

knowledge. In International Workshop on Business

Process Modeling, Development and Support (pp. 411-

425). Berlin, Heidelberg: Springer Berlin Heidelberg.

Sedrakyan, G., & Snoeck, M. (2017). Cognitive feedback

and behavioral feedforward automation perspectives

for modeling and validation in a learning context. In

Model-Driven Engineering and Software

Development: 4th International Conference,

MODELSWARD 2016, Rome, Italy, February 19-21,

2016, Revised Selected Papers 4 (pp. 70-92). Springer

International Publishing.

Sedrakyan, G., Poelmans, S., & Snoeck, M. (2017).

Assessing the influence of feedback-inclusive rapid

prototyping on understanding the semantics of parallel

UML statecharts by novice modellers. Information and

Software Technology, 82, 159-172.

Ruiz, J., Sedrakyan, G., & Snoeck, M. (2015, September).

Generating user interface from conceptual, presentation

and user models with JMermaid in a learning approach.

In Proceedings of the XVI International Conference on

Human Computer Interaction (pp. 1-8).

Sedrakyan, G., & Snoeck, M. (2014). Do we need to teach

testing skills in courses on requirements engineering

and modelling. In CEUR workshop proceedings (Vol.

1217, pp. 40-44). CEUR-WS. org.

Sedrakyan, G., Abdi, A., Van Den Berg, S. M., Veldkamp,

B. P., & Van Hillegersberg, J. (2022). Text-to-model

(tetomo) transformation framework to support

requirements analysis and modeling. In 10th

International Conference on Model-Driven

Engineering and Software Development,

MODELSWARD 2022 (pp. 129-136). SCITEPRESS.

OpenAI. (2024). ChatGPT (Nov 2024 version) Large

Language Model. OpenAI. https://chat.openai.com/

Sveidqvist, K., & Jain, A. (2021). The official guide to

Mermaid. js: create complex diagrams and beautiful

flowcharts easily using text and code. Packt Publishing.

Murray, A., & Scuotto, V. (2015). The business model

canvas. Symphonya. Emerging Issues in Management,

94-109.

Cai, D., Wang, Y., Liu, L., & Shi, S. (2022, July). Recent

advances in retrieval-augmented text generation.

In Proceedings of the 45th international ACM SIGIR

conference on research and development in

information retrieval (pp. 3417-3419).

Ollama. (n.d.). Ollama Library. Retrieved November 30,

2024, from https://ollama.com/library

Rodríguez, S. (n.d.). FastAPI documentation. FastAPI.

Retrieved November 30, 2024, from

https://fastapi.tiangolo.com/

Sedrakyan, G., & Snoeck, M. (2016). Enriching Model

Execution with Feedback to Support Testing of

Semantic Conformance between Models and

Requirements.

Sedrakyan, G., Järvelä, S., & Kirschner, P. (2016).

Conceptual framework for feedback automation and

personalization for designing learning analytics

dashboards. In Conference EARLI SIG (Vol. 27).

Reinventing Low-Code: Value-Driven and Learning-Oriented Low-Code Development with SLLM-Integrated Approach

431