AI Workflow
Full overhaul of system navigation
The Challenge
As part of our major report builder project, our app teams needed to create report facades from scratch, which were more or less the code that defined the most important content to the front end. For example, someone working in Payables might need fields like Vendor Name, Invoice Number, and Posting Date with a filter that only showed invoices on hold. Each of those items would also need a good description to assist in understanding what it provided. Unfortunately, our developers were stretched thin, so the burden fell on my team and me to get it done before shipping in 6 months.
My Approach
I would be working with our proprietary programming language in the Kiro IDE to try and pull out the relevant content. There was business logic code and UI code for each business class, and the first step was to rely on the UI code to identify key items. After all, if it was surfaced in the interface, it was likely important for a report.
The Process
By having the model compare the two files, I was able to get solid output. It slowly started to get better and better as I refined the steering file to ensure it was finding everything. It got to the stage where it would verify if the content was utilized in customers’ tenants before adding an item, and then it would check the generated facade file against a user persona to determine if the results were actually valid.
The Results
By the end, the workflow generated descriptions for over 100 fields across Payables and Projects, cutting a manual process that normally took a week down to a few hours. It was a low-stakes way to prove LLMs could handle domain-specific writing, and it gave our UI descriptions consistency across dozens of separate business classes.