Audical

Case Study: Audical - Fraudulent Phone Call Detection Automation

Project Duration

June 2019 - August 2019

Roles

Software Engineer, Designer

Problem Statement

Detecting fraudulent phone calls manually is a time-consuming and inefficient process. Analysts are required to listen to entire calls and search for suspicious behavior, leading to significant resource investment. There is a need for a desktop application that automates this process using AI and a weight-based algorithm. The application should simplify call analysis, automate speech-to-text conversion, and enable keyword searches to efficiently detect fraudulent phone calls.
Project Image

Introduction

As technology advances, so do the measures required to maintain security. Phone fraud has become increasingly prevalent, with the number of fraud attempts steadily increasing each year. Major players in the financial sector need to efficiently detect fraudulent phone calls to protect their clients. Our client, Booz Allen Hamilton, had an existing manual approach where an analyst would listen to a full phone call and look for any suspicious behavior. My challenge was to automate this process by designing and developing a desktop application.

Discover

During the discovery stage, I conducted user research to help with the planning phase. The key insights that helped define the final version of our product were that the transcription functionality was critical, and listening to an entire call was time-consuming, taking 30-45 minutes.

Define

To develop a clear objective, I brainstormed features for the product roadmap. I created a diagram that organized the main concerns of an analyst, which included simplifying the process of listening to calls, automating the process of listening to calls using speech-to-text AI, and searching for certain keywords and phrases that may indicate suspicious behavior.
Wireframe

Ideation

With enough information gathered, I created wireframes and a user flow to help diagram how a user might use the app, organizing the way I laid out the code.

Feedback

The wireframe layout was well received, but analysts suggested making phone printing and voice printing optional steps, analyzing multiple files at once, and having a separate window to indicate what files had been analyzed and which were remaining.

Development and Delivery

With feedback in mind, I created a high-fidelity model that was ultimately what the final product looked like. The application had three options: Single File Analyzer, Batch Analyzer, and Open Database.

The Single File Analyzer had two options that let an analyst Phoneprint, Voiceprint, or perform both. By giving analysts this freedom, they could omit extra metadata from being included in the analysis. If an analyst clicks “Phoneprint”, then only the “Enter Phone Number” field would pop-up. If an analyst also clicks “Voiceprint”, then they would have the option of uploading an audio file as well. Upon clicking Analyze, another window will pop up and indicate which files have been analyzed and which ones are still in the process.

The Batch File Analyzer performs similarly to the Single File Analyzer, but only allows for voice printing. It scans all the audio files within a folder, writes the metadata and attributes of the call to the Excel file, and then displays a popup message indicating that the analysis has been completed.

The Open Database feature was added to make it easier for analysts to open up the Excel spreadsheet with analyzed data written to it.

Conclusion

Creating a desktop application in two months was no easy task, but the final product successfully automated the process of detecting fraudulent phone calls using AI and a weight-based algorithm.