Fyber Hackathon – Part II – The Winners

The winners of Fyber Engineering’s 2020 Hackathon were Team RecogB-ad!

This team developed a POC for an innovative solution to a cross-industry problem.  The team was made up of members from both the Fyber Israel and the Berlin offices and included representatives from:

 

  • Data Science (Daniel Hen, Michael Winer)
  • Backend (Noy Miran, Marcelo Pereira)
  • DevOps (Ori Bracha).

 

Problem and Motivation

The problem the group wanted to find a solution to was with Publisher Blindness

 

Many of you may wonder, what does this mean?  

 

Using our Fyber Marketplace, publishers ask for ads from our integrated DSPs. However, they are unable to see the real content of those ads. An advertiser can publish whatever content they want. (Disclaimer: There is a necessary ethics code to follow). 

 

With this data at hand, publishers will be more confident and informed, thereby enabling them to make optimal decisions about advertisements presented within their apps.

 

The Solution

The solution that the team developed relies on three primary concepts:

  • Data extraction, processing and enrichment
  • AI Service
  • Publisher UI tools

The data which the team extracted is the advertiser’s campaigns, including other beneficial data.


The data which the team extracted is the advertiser’s campaigns, including their
iurl field.

 

The team fetched a sample of this data using Fyber’s internal tools.  In the future, there will be an automatic sampling process, using Kafka and Spark for this Big-Data task.

 

Once the team had retrieved the data, they could process and enrich this information.  Meaning, the publisher gains much more information and insights, based on different parameters.

 

The team was then able to use AWS Rekognition, which acted as their AI service, as part of the POC.  This service can perform relevant segmentation, classification and many more tasks which the team found relevant as part of this unique use case.

 

AWS Rekognition was able to generate relevant categories from each given image.  In the future, the team intends to develop an in-house solution to solve similar problems.

 

The end goal was to create a tool for the publisher to gain insights from these “discoveries”.

 

Our thoughts for the production architecture

 

This will allow our publishers to flag irrelevant categories and to block the advertisements accordingly.

 

Using Apache Druid, which is a high-performance real-time analytics database, the team managed to create a dedicated publisher UI, which contained, for example, ad category impressions breakdown, based on the categories recognized by the AI service.

 

The team also created another UI tool, which acted as the “Publisher Admin Console.”  The dashboard will allow the publisher to see statistics about the creatives, block “inappropriate” content and much more.

 

These products are expected to be integrated into Fyber’s platform, providing a complete experience, with a wider vision to the Fyber’s clients.

Read these next