Industries
Solution
Consulting
TCS Consulting Services to Achieve Success in the Digital Era
Highlights
- Many BFSI organizations still work with complex mainframe-based applications that, while reliable, are costly to maintain and operate, and don’t offer the kind of agility and resilience needed in today’s dynamic business environment.
- The industry realizes the pressing need to modernize mainframe systems, but the inherent complexities in reverse engineering and forward engineering legacy applications pose a major hurdle.
- Newer versions of large language models (LLMs), increased context token size, and retrieval augmented generation (RAG) capabilities will enable financial services firms to tackle the challenges in mainframe modernization effectively.
On this Page
Mainframe modernization an age-old problem
The core systems of all the major banking, financial services, and insurance (BFSI) companies will inherently have some legacy components written in old programming languages such as COBOL and Assembler.
The components include complex mainframe-based applications that are exceptionally reliable. However, the cost involved in their maintenance and operations adds up to of a fortune. These legacy applications are also notorious for their lack of agility. It takes a lot of time and money for organizations to add new business capabilities to these systems, create better user journeys, or respond to regulatory mandates.
Even though most of these enterprises desire to modernize their mainframe applications, the inherent complexities in reverse engineering and forward engineering the existing legacy applications pose a major hurdle. We believe GenAI can help.
With the advent of newer versions of large language models (LLMs), increased context token size, and retrieval-augmented generation (RAG) capabilities, GenAI is well poised to help enterprises tackle the challenges in mainframe modernization programs.
The GenAI advantage
Traditionally, automated tools for application discovery have been devised to analyze the legacy estate and generate dependency reports and other insights.
Moreover, there are a variety of code refactoring tools with various levels of automated code generation capabilities. But how GenAI-based tools are revolutionizing this area by providing high-quality output be it in business rules extraction (BRE) or target code generation with microservices architecture compatibility. For example, some of the newer GenAI based tools can analyze mainframe code and explain what it does in plain English and then generate corresponding Java code, enabling the same functionality with minimal inputs from the subject matter experts (SMEs). Once the BRE is done for a piece of legacy code, there is also the option of generating the target code based on those rules using GenAI tools like GitHub Copilot or Amazon CodeWhisperer. Initial pilots have demonstrated that the GenAI-generated code is of better quality and maintainability compared with traditional code refactoring approaches. This holds great promise as a modernization accelerator.
Challenges and mitigation options
Just like any other evolving technology, GenAI has its own challenges.
Due to LLMs’ inherent nature of being non-deterministic, GenAI output can vary over time for the same input prompt. It can also hallucinate and generate non-relevant output as we have experienced with the likes of ChatGPT. Yet another challenge is the limited context tokens that can be accepted by the LLMs to create and maintain the context of the prompts. There is also another set of challenges from a regulatory standpoint around the explainability, traceability, ownership, and nonrepudiation aspects of the generated output.
The newer versions of GenAI tools show a lot of promise in tackling the above-mentioned challenges by refining their architecture and by introducing innovative capabilities like RAG and multiple interpolation techniques. The RAG capability will help in reducing the effects of the LLMs hallucinations and non-deterministic behavior since it allows the GenAI solution architects to inject context to the GenAI tool. This can be achieved by vectorizing domain-related or context-setting data and feeding it to the GenAI solution to bind its output within the given context, resulting in a more accurate and relevant output that makes sense to the business area in question.
To increase the context length, a myriad techniques are available, including neural tangent kernel (NTK) interpolation, position interpolation (PI), attention with linear biases (ALiBi), and rotational positional encoding (RoPE) among others. These approaches have the potential to increase the context length to two million tokens and beyond. Any increase in the context length will have positive impact on the quality of output of the GenAI tools and solutions. For example, consider using GenAI to extract business rules from legacy application code and with larger context length. A greater number of legacy code files can be ingested in a single go, thereby enabling it to generate well-rounded business rules for a particular code module.
Newer LLMs will increase the success rate of modernization initiatives
A modernization road map with a well-defined GenAI strategy using RAG capabilities will be pivotal in accelerating any large modernization programs within the BFSI landscape.
It will increase the success rate of the mainframe optimization and modernization endeavours. With newer versions of LLMs and the faster rate of GenAI innovations, the future looks quite bright and promising for GenAI-enabled mainframe modernization initiatives. It would be advisable to start early experiments and pilots with the latest GenAI tools to understand their potential and experience their capabilities in accelerating the modernization programs.
