Summary
This interface was designed to be used by internal employees to train AI Model
Role
User Experience Lead
Company
Grainger
Problem
At the inception of the reverse image search machine learning model (see
Photo Search); we observed that the results it returned were inaccurate, causing users to quickly lose trust in the feature.
This
We attributed this to the limited diversity of training images for the model.
How might we improve photo search to make it more accurate and trustworthy? We realized that, to make the photo search feature useful for the user, we needed to improve its accuracy by training the model with uncurated images.
Customers send our customer support teams ~125k uncurated images every year.
Solution Hypothesis
If the customer support teams can use an interface to perform a photo search with uncurated images and provide accuracy feedback to the model, we should be able to improve its accuracy.
We set out to design an interface for customer support associates to use
Userflow
The userflow takes into account the customer support teams primary task which is to help the customer and focuses on optimizing time on task.
Wireframes
Wireframes show the image upload, crop and accuracy feedback screens
Result
The uncurated images and feedback data collected from customer support associates via this interface helped improve accuracy from 64% to 87%
We also observed an avg 2 min reduction in customer handle time for queries that included images as a result of using this photo search interface.
