Modernizing legacy applications with generative AI: Lessons from R&D Projects

As digital transformation accelerates, modernizing legacy applications has become essential for businesses to stay competitive. The application modernization market size, valued at  USD 21.32 billion in 2023 , is projected to reach  USD 74.63 billion by 2031 (1), reflecting the growing importance of updating outdated systems.

With 94% of business executives viewing AI as key to future success and 76% increasing their investments in Generative AI due to its proven value (2), it's clear that AI is becoming a critical driver of innovation. One key area where AI is making a significant impact is  application modernization - an essential step for businesses aiming to improve scalability, performance, and efficiency.

Based on  two projects conducted by our  R&D team , we've seen firsthand how Generative AI can streamline the process of rewriting legacy systems.

Let’s start by discussing the importance of  rewriting legacy systems and how GenAI-driven solutions are transforming this process.

Why re-write applications?

In the rapidly evolving software development landscape, keeping applications up-to-date with the latest programming languages and technologies is crucial. Rewriting applications to new languages and frameworks can significantly enhance performance, security, and maintainability. However, this process is often labor-intensive and prone to human error.

 Generative AI offers a transformative approach to code translation by:

•  leveraging advanced machine learning models to automate the rewriting process
•  ensuring consistency and efficiency
•  accelerating modernization of legacy systems
•  facilitating cross-platform development and code refactoring

As businesses strive to stay competitive, adopting Generative AI for code translation becomes increasingly important. It enables them to harness the full potential of modern technologies while minimizing risks associated with manual rewrites.

Legacy systems, often built on outdated technologies, pose significant challenges in terms of maintenance and scalability. Modernizing legacy applications with Generative AI provides a viable solution for rewriting these systems into modern programming languages, thereby extending their lifespan and improving their integration with contemporary software ecosystems.

This automated approach not only preserves core functionality but also enhances performance and security, making it easier for organizations to adapt to changing technological landscapes without the need for extensive manual intervention.

Why Generative AI?

Generative AI offers a powerful solution for rewriting applications, providing several key benefits that streamline the modernization process.

Modernizing legacy applications with Generative AI proves especially beneficial in this context for the following reasons:

•     Identifying relationships and business rules:    Generative AI can analyze legacy code to uncover complex dependencies and embedded business rules, ensuring critical functionalities are preserved and enhanced in the new system.
•     Enhanced accuracy:    Automating tasks like code analysis and documentation, Generative AI reduces human errors and ensures precise translation of legacy functionalities, resulting in a more reliable application.
•     Reduced development time and cost:    Automation significantly cuts down the time and resources needed for rewriting systems. Faster development cycles and fewer human hours required for coding and testing lower the overall project cost.
•     Improved security:    Generative AI aids in implementing advanced security measures in the new system, reducing the risk of threats and identifying vulnerabilities, which is crucial for modern applications.
•     Performance optimization:    Generative AI enables the creation of optimized code from the start, integrating advanced algorithms that improve efficiency and adaptability, often missing in older systems.

By leveraging Generative AI, organizations can achieve a smooth transition to modern system architectures, ensuring substantial returns in performance, scalability, and maintenance costs.

In this article, we will explore:

•  the use of Generative AI for rewriting a simple CRUD application
•  the use of Generative AI for rewriting a microservice-based application
•  the challenges associated with using Generative AI

For these case studies, we used OpenAI's ChatGPT-4 with a context of 32k tokens to automate the rewriting process, demonstrating its advanced capabilities in understanding and generating code across different application architectures.

We'll also present the benefits of using  a data analytics platform designed by Grape Up's experts. The platform utilizes Generative AI and neural graphs to enhance its data analysis capabilities, particularly in data integration, analytics, visualization, and insights automation.

Project 1: Simple CRUD application

The  source CRUD project was used as an example of a simple CRUD application - one written utilizing .Net Core as a framework, Entity Framework Core for the ORM, and SQL Server for a relational database. The target project containes a backend application created using Java 17 and Spring Boot 3.

Steps taken to conclude the project

Rewriting a simple CRUD application using Generative AI involves a series of methodical steps to ensure a smooth transition from the old codebase to the new one. Below are the key actions undertaken during this process:

•     initial architecture and data flow investigation    - conducting a thorough analysis of the existing application's architecture and data flow.
•     generating target application skeleton    - creating the initial skeleton of the new application in the target language and framework.
•     converting components    - translating individual components from the original codebase to the new environment, ensuring that all CRUD operations were accurately replicated.
•     generating tests    - creating automated tests for the backend to ensure functionality and reliability.

Throughout each step, some manual intervention by developers was required to address code errors, compilation issues, and other problems encountered after using OpenAI's tools.

Initial architecture and data flows’ investigation

The first stage in rewriting a simple CRUD application using Generative AI is to conduct a thorough investigation of the existing architecture and data flow. This foundational step is crucial for understanding the current system's structure, dependencies, and business logic.

This involved:

•     codebase analysis  
•     data flow mapping    – from user inputs to database operations and back
•     dependency identification  
•     business logic extraction    – documenting the core business logic embedded within the application

While  OpenAI's ChatGPT-4 is powerful, it has some limitations when dealing with large inputs or generating comprehensive explanations of entire projects. For example:

•  OpenAI couldn’t read files directly from the file system
•  Inputting several project files at once often resulted in unclear or overly general outputs

However, OpenAI excels at explaining large pieces of code or individual components. This capability aids in understanding the responsibilities of different components and their data flows. Despite this, developers had to conduct detailed investigations and analyses manually to ensure a complete and accurate understanding of the existing system.

This is the point at which we used our data analytics platform. In comparison to OpenAI, it focuses on data analysis. It's especially useful for analyzing data flows and project architecture, particularly thanks to its ability to process and visualize complex datasets. While it does not directly analyze source code, it can provide valuable insights into how data moves through a system and how different components interact.

Moreover, the platform excels at visualizing and analyzing data flows within your application. This can help identify inefficiencies, bottlenecks, and opportunities for optimization in the architecture.

Generating target application skeleton

As with OpenAI's inability to analyze the entire project, the attempt to generate the skeleton of the target application was also unsuccessful, so the developer had to manually create it. To facilitate this,  Spring Initializr was used with the following configuration:

•  Java: 17
•  Spring Boot: 3.2.2
•  Gradle: 8.5

Attempts to query OpenAI for the necessary Spring dependencies faced challenges due to significant differences between dependencies for C# and Java projects. Consequently, all required dependencies were added manually.

Additionally, the project included a database setup. While OpenAI provided a series of steps for adding database configuration to a Spring Boot application, these steps needed to be verified and implemented manually.

Converting components

After setting up the backend, the next step involved converting all project files - Controllers, Services, and Data Access layers - from C# to Java Spring Boot using OpenAI.

The AI proved effective in converting endpoints and data access layers, producing accurate translations with only minor errors, such as misspelled function names or calls to non-existent functions.

In cases where non-existent functions were generated, OpenAI was able to create the function bodies based on prompts describing their intended functionality. Additionally, OpenAI efficiently generated documentation for classes and functions.

However, it faced challenges when converting components with extensive framework-specific code. Due to differences between frameworks in various languages, the AI sometimes lost context and produced unusable code.

Overall, OpenAI excelled at:

•  converting data access components
•  generating REST APIs

However, it struggled with:

•  service-layer components
•  framework-specific code where direct mapping between programming languages was not possible

Despite these limitations, OpenAI significantly accelerated the conversion process, although manual intervention was required to address specific issues and ensure high-quality code.

Generating tests

Generating tests for the new code is a crucial step in ensuring the reliability and correctness of the rewritten application. This involves creating both  unit tests and  integration tests to validate individual components and their interactions within the system.

To create a new test, the entire component code was passed to OpenAI with the query:  "Write Spring Boot test class for selected code."

OpenAI performed well at generating both integration tests and unit tests; however, there were some distinctions:

•     For unit tests    , OpenAI generated a new test for each if-clause in the method under test by default.
•     For integration tests    , only happy-path scenarios were generated with the given query.
•     Error scenarios    could also be generated by OpenAI, but these required more manual fixes due to a higher number of code issues.

If the test name is self-descriptive, OpenAI was able to generate unit tests with a lower number of errors.

Project 2: Microservice-based application

As an example of a microservice-based application, we used the  Source microservice project - an application built using .Net Core as the framework, Entity Framework Core for the ORM, and a Command Query Responsibility Segregation (CQRS) approach for managing and querying entities.  RabbitMQ was used to implement the CQRS approach and  EventStore to store events and entity objects. Each microservice could be built using Docker, with  docker-compose managing the dependencies between microservices and running them together.

The target project includes:

•  a microservice-based backend application created with     Java 17    and     Spring Boot 3  
•  a frontend application using the     React    framework
•     Docker support    for each microservice
•     docker-compose    to run all microservices at once

Project stages

Similarly to the CRUD application rewriting project, converting a microservice-based application using Generative AI requires a series of steps to ensure a seamless transition from the old codebase to the new one. Below are the key steps undertaken during this process:

•     initial architecture and data flows’ investigation    - conducting a thorough analysis of the existing application's architecture and data flow.
•     rewriting backend microservices    - selecting an appropriate framework for implementing CQRS in Java, setting up a microservice skeleton, and translating the core business logic from the original language to Java Spring Boot.
•     generating a new frontend application    - developing a new frontend application using React to communicate with the backend microservices via REST APIs.
•     generating tests for the frontend application    - creating unit tests and integration tests to validate its functionality and interactions with the backend.
•     containerizing new applications    - generating Docker files for each microservice and a docker-compose file to manage the deployment and orchestration of the entire application stack.

Throughout each step, developers were required to intervene manually to address code errors, compilation issues, and other problems encountered after using OpenAI's tools. This approach ensured that the new application retains the functionality and reliability of the original system while leveraging modern technologies and best practices.

Initial architecture and data flows’ investigation

The first step in converting a microservice-based application using Generative AI is to conduct a thorough investigation of the existing architecture and data flows. This foundational step is crucial for understanding:

•  the system’s structure
•  its dependencies
•  interactions between microservices

 Challenges with OpenAI
Similar to the process for a simple CRUD application, at the time, OpenAI struggled with larger inputs and failed to generate a comprehensive explanation of the entire project. Attempts to describe the project or its data flows were unsuccessful because inputting several project files at once often resulted in unclear and overly general outputs.

 OpenAI’s strengths
Despite these limitations, OpenAI proved effective in explaining large pieces of code or individual components. This capability helped in understanding:

•  the responsibilities of different components
•  their respective data flows

Developers can create a comprehensive blueprint for the new application by thoroughly investigating the initial architecture and data flows. This step ensures that all critical aspects of the existing system are understood and accounted for, paving the way for a successful transition to a modern microservice-based architecture using Generative AI.

Again, our data analytics platform was used in project architecture analysis. By identifying integration points between different application components, the platform helps ensure that the new application maintains necessary connections and data exchanges.

It can also provide a comprehensive view of your current architecture, highlighting interactions between different modules and services. This aids in planning the new architecture for efficiency and scalability. Furthermore, the platform's analytics capabilities support identifying potential risks in the rewriting process.

Rewriting backend microservices

Rewriting the backend of a microservice-based application involves several intricate steps, especially when working with specific architectural patterns like  CQRS (Command Query Responsibility Segregation) and  event sourcing . The source C# project uses the CQRS approach, implemented with frameworks such as  NServiceBus and  Aggregates , which facilitate message handling and event sourcing in the .NET ecosystem.

 Challenges with OpenAI
Unfortunately, OpenAI struggled with converting framework-specific logic from C# to Java. When asked to convert components using NServiceBus, OpenAI responded:

 "The provided C# code is using NServiceBus, a service bus for .NET, to handle messages. In Java Spring Boot, we don't have an exact equivalent of NServiceBus, but here's how you might convert the given C# code to Java Spring Boot..."

However, the generated code did not adequately cover the CQRS approach or event-sourcing mechanisms.

 Choosing Axon framework
Due to these limitations, developers needed to investigate suitable Java frameworks. After thorough research, the     Axon Framework   was selected, as it offers comprehensive support for:

•     domain-driven design  
•     CQRS  
•     event sourcing  

Moreover, Axon provides out-of-the-box solutions for message brokering and event handling and has a  Spring Boot integration library , making it a popular choice for building Java microservices based on CQRS.

 Converting microservices
Each microservice from the source project could be converted to  Java Spring Boot using a systematic approach, similar to converting a simple CRUD application. The process included:

•  analyzing the data flow within each microservice to understand interactions and dependencies
•  using        Spring Initializr      to create the initial skeleton for each microservice
•  translating the core business logic, API endpoints, and data access layers from C# to Java
•  creating unit and integration tests to validate each microservice’s functionality
•  setting up the event sourcing mechanism and CQRS using the Axon Framework, including configuring Axon components and repositories for event sourcing

 Manual Intervention
Due to the lack of direct mapping between the source project's CQRS framework and the Axon Framework, manual intervention was necessary. Developers had to implement framework-specific logic manually to ensure the new system retained the original's functionality and reliability.

Generating a new frontend application

The source project included a frontend component written using  aspnetcore-https and  aspnetcore-react libraries, allowing for the development of frontend components in both C# and React.

However, OpenAI struggled to convert this mixed codebase into a React-only application due to the extensive use of C#.

Consequently, it proved faster and more efficient to generate a new frontend application from scratch, leveraging the existing REST endpoints on the backend.

Similar to the process for a simple CRUD application, when prompted with  “Generate React application which is calling a given endpoint” , OpenAI provided a series of steps to create a React application from a template and offered sample code for the frontend.

•  OpenAI successfully generated React components for each endpoint
•  The CSS files from the source project were reusable in the new frontend to maintain the same styling of the web application.
•  However, the overall structure and architecture of the frontend application remained the developer's responsibility.

Despite its capabilities, OpenAI-generated components often exhibited issues such as:

•  mixing up code from different React versions, leading to code failures.
•  infinite rendering loops.

Additionally, there were challenges related to CORS policy and web security:

•  OpenAI could not resolve CORS issues autonomously but provided explanations and possible steps for configuring CORS policies on both the backend and frontend
•  It was unable to configure web security correctly.
•  Moreover, since web security involves configurations on the frontend and multiple backend services, OpenAI could only suggest common patterns and approaches for handling these cases, which ultimately required manual intervention.

Generating tests for the frontend application

Once the frontend components were completed, the next task was to generate tests for these components.  OpenAI proved to be quite effective in this area. When provided with the component code, OpenAI could generate simple unit tests using the  Jest library.

OpenAI was also capable of generating integration tests for the frontend application, which are crucial for verifying that different components work together as expected and that the application interacts correctly with backend services.

However, some  manual intervention was required to fix issues in the generated test code. The common problems encountered included:

•  mixing up code from different React versions, leading to code failures.
•  dependencies management conflicts, such as mixing up code from different test libraries.

Containerizing new application

The source application contained  Dockerfiles that built images for C# applications. OpenAI successfully converted these Dockerfiles to a new approach using  Java 17 ,  Spring Boot , and  Gradle build tools by responding to the query:

 "Could you convert selected code to run the same application but written in Java 17 Spring Boot with Gradle and Docker?"

Some manual updates, however, were needed to fix the actual jar name and file paths.

Once the React frontend application was implemented, OpenAI was able to generate a Dockerfile by responding to the query:

 "How to dockerize a React application?"

Still, manual fixes were required to:

•  replace paths to files and folders
•  correct mistakes that emerged when generating     multi-staged Dockerfiles    , requiring further adjustments

While OpenAI was effective in converting individual Dockerfiles, it struggled with writing  docker-compose files due to a lack of context regarding all services and their dependencies.

For instance, some microservices depend on database services, and OpenAI could not fully understand these relationships. As a result, the docker-compose file required significant manual intervention.

Conclusion

Modern tools like OpenAI's ChatGPT can significantly enhance software development productivity by automating various aspects of code writing and problem-solving. Leveraging large language models, such as OpenAI over ChatGPT can help generate large pieces of code, solve problems, and streamline certain tasks.

However, for complex projects based on microservices and specialized frameworks, developers still need to do considerable work manually, particularly in areas related to architecture, framework selection, and framework-specific code writing.

 What Generative AI is good at:

•     converting pieces of code from one language to another    - Generative AI  excels at translating individual code snippets between different programming languages, making it easier to migrate specific functionalities.
•     generating large pieces of new code from scratch    - OpenAI can generate substantial portions of new code, providing a solid foundation for further development.
•     generating unit and integration tests    - OpenAI is proficient in creating unit tests and integration tests, which are essential for validating the application's functionality and reliability.
•     describing what code does    - Generative AI can effectively explain the purpose and functionality of given code snippets, aiding in understanding and documentation.
•     investigating code issues and proposing possible solutions    - Generative AI can quickly analyze code issues and suggest potential fixes, speeding up the debugging process.
•     containerizing application    - OpenAI can create Dockerfiles for containerizing applications, facilitating consistent deployment environments.

At the time of project implementation,  Generative AI still had several limitations .

•  OpenAI struggled to provide comprehensive descriptions of an application's overall architecture and data flow, which are crucial for understanding complex systems.
•  It also had difficulty identifying equivalent frameworks when migrating applications, requiring developers to conduct manual research.
•  Setting up the foundational structure for microservices and configuring databases were tasks that still required significant developer intervention.
•  Additionally, OpenAI struggled with managing dependencies, configuring web security (including CORS policies), and establishing a proper project structure, often needing manual adjustments to ensure functionality.

Benefits of using the data analytics platform:

•     data flow visualization:    It provides detailed visualizations of data movement within applications, helping to map out critical pathways and dependencies that need attention during re-writing.
•     architectural insights    : The platform offers a comprehensive analysis of system architecture, identifying interactions between components to aid in designing an efficient new structure.
•     integration mapping:    It highlights integration points with other systems or components, ensuring that necessary integrations are maintained in the re-written application.
•     risk assessment:    The platform's analytics capabilities help identify potential risks in the transition process, allowing for proactive management and mitigation.

By leveraging GenerativeAI’s strengths and addressing its limitations through manual intervention, developers can achieve a more efficient and accurate transition to modern programming languages and technologies. This hybrid approach to modernizing legacy applications with Generative AI currently ensures that the new application retains the functionality and reliability of the original system while benefiting from the advancements in modern software development practices.

It's worth remembering that Generative AI technologies are rapidly advancing, with improvements in processing capabilities. As Generative AI  becomes more powerful, it is increasingly able to understand and manage complex project architectures and data flows. This evolution suggests that in the future, it will play a pivotal role in rewriting projects.

Do you need support in modernizing your legacy systems with expert-driven solutions?

.................

Sources:

1.  https://www.verifiedmarketresearch.com/product/application-modernization-market/
2.  https://www2.deloitte.com/content/dam/Deloitte/us/Documents/deloitte-analytics/us-ai-institute-state-of-ai-fifth-edition.pdf