icon 1

90%
of hateful comments
detected by the application

Icon 2

2% 
error margin
for the algorithm (false positives)

icon 3

2 million+ 
hateful comments 
deleted in 20 months

The background

Charles Cohen started learning about computer programming when he was just 10 years old. 11 years later, he launched his first mobile application — Bodyguard. Behind a seemingly simple idea, there was a complex goal — to protect web users with real-time protection against cyber-bullying.

Why did he choose to take on this challenge? The answer is simple — applications like this were not available then. Moderation did not prove to be efficient enough on other platforms.

“I have never been a victim of cyber-bullying myself, but I grew up with everyone around me using social networks. I could see the harm caused by online hate. Cyber-bullying restricts freedom of expression, and this is what I struggled with when I was a teenager. It’s why I never dared to create profiles online or expose myself on public platforms — I was fearful of being bullied.”

Charles Cohen, founder and CEO of Bodyguard

The challenge

The technology needed to be capable of analysing the context in which a comment is made, and determining the person or people it is aimed at.

Bodyguard technology had to be able to understand and interpret states of mind. For this reason, an artificial intelligence layer was absolutely vital to reduce false positives (comments detected as hateful when they are not), and increase accuracy.

"The technology also needed to be capable of detecting irony, sarcasm, and even humour. The predictive model created with the OVHcloud AutoML machine learning platform really helped me achieve this.”

Charles Cohen, founder and CEO of Bodyguard

The predictive model also needed to make the technology take into account the relationship between two individuals. For example, is the author of the comment ‘following’ the person they are replying to? This required research, and the exchange of nearly 80 pieces of metadata. In this data, the reaction time after publication would appear, along with the percentage of upper-case letters and the profile picture.

For this innovative project, Charles also needed to find a good enough algorithm among those offered by scikit-learn. This open-source library offers algorithms that are mostly written in Python, and designed for machine learning.

In terms of the specifications, it was a matter of accuracy. The predictive model could not exceed an error rate of 10%.

The solution

A managed, easy-to-use service that could accelerate the production phase.

The software layer

Charles chose to use OVHcloud AutoML, a distributed and scalable machine learning platform. With this Software-as-a-Service (SaaS) solution, he could automate the creation and deployment processes, as well as the process for requesting machine learning models. He could also use it to integrate open-source algorithms, such as those offered by scikit-learn.

OVHcloud AutoML also accelerated the development phase. 10 days were needed to create the Bodyguard predictive model, and it took 20 days to design the meta-learning model, which analyses the relationship between the author of the content and the commenter.

With these models, the detection accuracy of Bodyguard technology has increased by 10%, moving from 80% to 90%. The number of false positives has also decreased by 15%, from 6% to 3%.

In terms of monitoring, Charles chose to use the Logs Data Platform (currently available in France only) with Grafana software. With this technology, he can monitor the performance of his infrastructure and databases. It can also be used to measure key performance indicators (KPIs), i.e. the number of users, the volume of hateful comments deleted in real time, the number of API requests, and more.

The hardware layer

The Bodyguard infrastructure is made up of three Public Cloud instances:

  • one for managing the databases
  • another for the technology and machine learning models
  • another for the systems that keep the mobile application running, by gathering the comments and analysing them using the technology

And to create backups, Charles uses another service from the OVHcloud Public Cloud: Cloud Archive (currently available in France only). He can use it for long-term data storage at a lower cost, but still guarantee security and data recovery.

Bodyguard infrastructure 2

 

The result

Charles spent two years developing the final machine learning algorithm and integrating it into a free mobile application, which has been available on Android and iOS since October 2017. Today, Bodyguard deletes hateful comments in real time on YouTube, Instagram, Twitter, Twitch and Mixer.

In July 2019, this virtual bodyguard gathered a following of more than 40,000 users, and boasted a satisfaction score of 97%. Here are a few reasons why it is so successful:

  • 90% of hateful comments detected by the application
  • 2% error margin (false positives)
  • 2 million+ hateful comments deleted in 20 months

The application will soon be translated into English and Spanish. A new solution called “Bodyguard for Families” will also be available soon, to alert parents immediately if their children are being cyber-bullied.

Over time, Charles hopes to emerge as a cloud provider for AI-powered automatic moderation solutions. To do this, he is offering access to his technology under the name “Bodyguard for enterprises”, via an API. It is aimed at those who would like to protect themselves, their users, their image, their reputation and their employees.

“For this, our platform designed for developers (developers.bodyguard.ai) is already available, so that everyone can use our technology.”

Charles Cohen, founder and CEO of Bodyguard