The client is a multinational firm that develops enterprise software for business process management, integration, and big data analytics.
The Challenge
The client – a global software firm specializing in B2B enterprise applications – needed to develop a more modern and efficient Integration and API Management Platform. However, data interchange rules are typically expressed in natural language by customers, so developers needed to “understand the rule” before creating a workflow tool using a drag-and-drop UI. If this is managed as a manual process, it inevitably involves many repetitive tasks, and the entire process is prone to human error.
The Solution
Persistent used a combination of Natural Language Processing (NLP) techniques and Generative AI with a “Few-Shot” training approach to understand the complex rules explained in natural language. Few-Shot training is a type of Machine Learning (ML) that seeks to classify new data when AI has only a few training samples from which to learn. Put another way, with Few-Shot learning, computers are expected to detect generalized patterns and rules from a few examples, the same way that humans do.
Our AI team crafted custom prompts for each rule type while providing sufficient flexibility in the solution to accommodate unseen rule expressions and minimizing Large Language Model (LLM) expense. We also created a “playground” to allow developers to access a rich library of sample formats, which were then consumed and processed by code transpilers to improve the accuracy of the system.
The entire solution employed OpenAI GPT 4 APIs, and benefitted from the scale and processing power of the Microsoft Azure cloud. The team also relied on LangChain as a language model integration framework, given that LangChain’s use cases largely overlap with language models in general, including document analysis and summarization.
The Outcome
The GenAI solution achieved an impressive 80% automation of rules, with a 60% savings in total FTE (labor) cost. The automation of the Integration and API Management process also inherently improves accuracy by avoiding the human errors associated with manual data management.
Technology used:
- Transpilers
- NLP
- MS Azure OpenAI GPT4 APIs
- LangChain
- Custom Prompt Engineering
