Natural language filtering
Make existing product capabilities easier to discover.
Calls from last week where the customer mentioned cancellation.
Problem
ACE’s call filters could answer useful questions, but users had to know which filters to look for and how to combine them. I proposed a natural language entry point so they could describe the view they needed.
My contribution
I led the work from early experiments through AI and backend implementation, product integration, deployment, and observability. Retrieval supplied context for organization-specific values, while checks against available values helped resolve what a request referred to.
Decision
After inspecting production requests, I added deterministic handling for recurring simple queries and kept model-assisted interpretation for requests that needed it. Real usage also informed evaluation examples and filter descriptions.
Outcome
The feature covered 40+ predefined filters, with filter activity up 13% in the first month. Activity means page and API requests, not unique users. About 40% of queries took a deterministic path, reducing latency for that subset from 5 seconds to under 1 second.