The problem
A small online media company that accepts user-generated content was being overwhelmed by the volume of content and the need to moderate it for anything inappropriate. It ran a small army of offshore moderators working 8-to-10-hour shifts, reviewing each piece of content by hand before it went live.
The work was monotonous, and accuracy dropped once reviewers hit a certain level of fatigue. The real exposure, though, was commercial: failure to adequately moderate the content could cost the publisher its merchant banking account — and with it, the ability to accept credit-card payments.
What we did
Reviewing every submission by hand was inherently unscalable. The Scagility team looked at reducing that load with machine vision — automating the first pass and flagging only the exceptions for a human to see.
Working with open-source models, we had a machine-vision prototype running in two weeks. It reduced human moderation volumes by 96% by surfacing the content most likely to be inappropriate, and its accuracy improved over time as reviewers validated its calls.