A leading fashion ecommerce company that uses AI-based recommendation systems wanted to find an easier way to tap into the knowledge of their domain experts and improve their algorithmically generated ads. 

One of the earliest data science teams to adopt Labelbox focused on innovating on personalized ads algorithmically. The team collaborates closely with the marketing creative team given that for their service, creative assets are a key generator of conversion, user adoption and sales revenue. As it is common to find in retail and ecommerce, the company had millions of potential unstructured images from which they could choose from in order to create these brand ads and campaigns. Showing professionally produced outfit matches would typically yield strong conversion, but quickly became a bottleneck because of the time and resources needed to produce them at scale. Instead, the team wanted to find ways to generate these creative assets algorithmically and pair them with the right copy and text. 

To solve this, the company utilized natural language processing (NLP) and computer vision for both creating the copy, as well as the images of outfit combinations that were being algorithmically generated. The Labelbox platform was used as the core infrastructure for annotation and collaboration in order to ensure that there was a human QA layer to review all of these text and visual assets. 

Key questions typically answered inside Labelbox included: “Would this layout be acceptable?” and “Would this outfit combination be acceptable?”, while tapping into subject matter experts for inputting their judgment and expertise. Labelbox provided the ability to essentially “quiz” these domain experts - stylists in this instance - to offer their recommendations and train ML models faster in order to deliver better predictions. In addition, the copywriters who were tasked with creating headlines for these ads benefited from using Labelbox as the centralized place for all their work and no longer having to store their answers in disparate places that are hard to track and reuse, such as clunky cloud-based spreadsheets. 

The efficiency gains from this new process of producing these personalized ads were striking. After doing a before and after comparison test, the data science team discovered that their prior process - which typically took 2 weeks for producing a set of ad assets - now just took a few days. Having a central platform to automate the data import and export process, speed up human QA review, and simplify the management of a high number of users allowed the company to scale this algorithmically-driven approach for creating personalized ads. 

Fast forward to today, the company now manages nearly a thousand users across multiple teams and roughly 90 projects. The company continues to produce more effective marketing campaigns because they have an easier way to systematize their data labeling process, while utilizing a single destination for building unique use cases for testing and productionizing AI.