KAI: An AI-powered web module for improving online and in-person communication

NERDZ LAB’s task was to develop an integrated web-based module as an additional internal component to the client’s AI solutions. The module had to integrate with any website, teach AI models to identify areas for improvement, and then provide sales reps with actionable outcomes. The module would help KAI’s team of in-house testers quickly and efficiently assess the quality of text-based interactions. 

Here’s how it works: 

• • The module converts a recorded conversation between a sales representative and a prospective client into text. 
  • Next, the tester listens to the audio and suggests how the conversation could be improved—which words to use, which style and tone of voice would be more appropriate, and so on. 
  • This feedback is then used to teach KAI’s AI algorithm for sales interaction training.

Our team identified the functionality that KAI’s AI learning software had to offer so that testers could optimize their workflow. There were four main challenges:

• A custom text editor. We had to create a custom text editor since the standard one didn’t support the required editing features. For example, text needed to be able to show graphical elements such as badges, highlights, and tags. Moreover, the text editor had to include other visual features—speakers—laid out as the editor image on the left with the name above the text. We had to make speaker elements a part of the general text editor so that testers could work with the text throughout the entire conversion, not just with one speaker’s part.
• Lengthy conversations. The average 30-minute call contains approximately 15,000 words. That’s a lot of data to process—and a common reason for a system freeze. Our solution had to load all this data while keeping real-time performance sharp since every second of a system freeze negatively impacts a tester’s work. A tester may need to change the text quickly and type without long pauses. Consequently, our task was to make the entire system run smoothly and give the user the feeling that they are simply typing, with no complex processes in the background.
• Text and audio synchronization. Each word has its own timecode, which allows the system to highlight words as they are played in the audio. However, when a tester makes changes to the text—such as adding or deleting words while the audio remains unchanged—these timecodes have to be recalculated. Timecode recalculation is challenging, as each word’s timecode depends on the timecodes of neighboring words, the length of the text, the way the word is pronounced, and so on. We had to take all of these things into account to ensure timecode accuracy.

• Search and edit issue. When a tester highlighted, tagged, or made any other visual change to the text, this would trigger a re-rendering of the entire text component. This created a significant system lag which was negatively impacting user experience. The most challenging task was to render only those text components that were highlighted, tagged, or otherwise changed without affecting the other components.