Otaku rolls out a subscription model thanks to a system for personalising content developed by Sciling

Manga and anime are no longer genres for just a small group of fans of Japanese culture, but have become a worldwide phenomenon with ever-more followers. This Eastern boom in the West has led to numerous websites being created that compete with each other for the biggest number of followers.


Otaku* is one such website. It is a platform that provides the latest comprehensive coverage of anime, manga, games and any other topic one can think of related to Japanese culture. Nevertheless, in the hyper-competitive universe of online platforms, the quality and variety of the content are no longer sufficient to ensure success for such a model.

Whatever we call it—the “Netflix” or “Amazon” effect—the fact is that in any kind of experience, consumers expect increasingly more hyper-personalisation of the kind they get from such technological giants. This is even more relevant with digital experiences. The creator of Otaku was aware that the success of their business model depended on being able to provide the best experience possible for an ever-more demanding kind of customer. Artificial Intelligence was to be fundamental in achieving that aim.

*Real name protected by non-disclosure contract

About the client

Otaku is a digital social platform for anime and manga lovers. Via their website or mobile app, the fans of these genres can get updates about the latest news, discover new artists, contribute content, take part in forums and discussions about Eastern culture and get to know people with the same interests.

The challenge

Otaku, like Spotify, works using a freemium model. In other words, advanced services are offered to users who pay a fee. Its creator understood that, just like its counterpart in the world of music, it was essential to know the users’ tastes better than anybody in order to give them the type of content they were looking for quickly and smoothly. With over 20 million publications of greatly varying kinds, personalisation was fundamental to ensure the users’ participation and retain them.

We could say that finding the desired content from among 20 million options is the modern equivalent of looking for a needle in a haystack. So, as soon as the platform was launched, a system was implemented to personalise the continent. However, just four years later, the system had become unstable. It was consuming 20 times more memory than at the start and needed a complex storage system to work.

After three failed attempts to replace it for a new, more reliable one, and in keeping with their current needs, the creator of Otaku had practically given up hope of finding a solution to the problem. “We had invested a lot of time and money; as much in attempting to create a new personalisation system as in keeping the initial one alive.

But all we had achieved was to make the system more complicated and unstable. The risk that everything would collapse like a house of cards and completely stop working was becoming ever-more real. I was desperate.”

And he had good reason to be. The problem was so great that by the time Sciling finished introducing the new personalisation system, the old one had become completely inoperative.

The requirements that the team at Sciling had to tackle were anything but simple. The new system would have to deal with a database of 950,000 users, a catalogue with 20 million publications and a rate of 25 requests per second with over 800 million registrations in the database. And all of this was running with a local machine of just 64 GB of RAM memory. “It was not going to be easy, but neither was it the first time we were facing such tight conditions,” explains Vicent Alabau, Director of Operations at Sciling.

“The team at Sciling enthused me with the necessary confidence to take on a fourth attempt. What’s more, the streamlined way they carry out their projects was one of the essential aspects to convince the rest of the management board that it was worth trying again. For me, it was vital for this project to be a success. After three failed attempts, I was taking a chance in front of my partners by proposing this last attempt. My credibility would have been very badly shaken if it didn’t work out. But with Sciling I was sure that the risk of taking on this new project was infinitely lower than the risk of being left without a service,” says the creator of Otaku.

Strategic vision

To optimise the system to the full, two optimisation methods were used: a computational one and an algorithmic one. On the one hand, the system was introduced through neural networks to take advantage of the superior performance of the graphic processing units (GPUs). On the other, a system for grouping articles was set up to speed up the internal search procedures.

Our solution

From gathering requirements to fully integrating with Otaku’s database, as well as analysing the different options, Sciling carried out the complete process of replacing the old recommendation system for a new, much more robust one which has also clearly improved the results obtained by the previous one. “There was such a difference that some users asked us how we did it. But the important thing is that we no longer had to worry about leaving them with no service because the old one had stopped working from one day to another,” explains the creator of Otaku.

The new system takes the utmost advantage of the client’s technological infrastructure by using each one of the four GPUs for a different task: one GPU for the item-item recommender, one for the item-user recommender, one for re-training the model, and another to fine-tune the model. Technology for grouping in clusters within the GPU was also used to ensure the 25 requests a second. “We came up with solutions so that the very limited technical specifications wouldn’t affect the project’s end result. The new system has been working for a year, and has not collapsed even once,” affirms Germán Sanchis, CEO of Sciling.

Launching this new recommendation system has enabled Otaku to ensure a constant, loyal premium customer base, thereby ensuring the visibility of their business model. In fact, personalisation of content is one of the best-rated characteristics of their subscription service, which today has accumulated 7,600 loyal users.

But not only has the number of users grown who consider the $10 a month they pay for the premium package to be more than fair given the value they receive. The other business metrics have also improved. “The time measured for visits has risen to ten and a half minutes. We are at Facebook levels!” they enthuse at Otaku’s analytics department.

Otaku is still working side-by-side with Sciling to continue to improve the users’ experience. In addition to showing them the most relevant publications, soon the order in which these publications are shown will be given by the user’s preferences.



20 Million


800 Million

Data points


Request per second

It is difficult to find talented people in the world of website development, and much more so in the sphere of machine learning. But Sciling’s team has proven to be very talented.


Wow, buddy! What have you done? That’s amazing!

Platform user

The swift approximation that Sciling uses was an essential factor. Their analysis of the initial risks and options helped me to reduce the management board’s uncertainty and get the green light from them for this project.


The implementation process

During the initial analysis, deep learning techniques stood out as the most efficient and scalable alternative. It was clear they had the potential to improve the quality of recommendations given to users compared to traditional techniques. On the one hand, they enable characteristics to be extracted very well, achieving greater accuracy in posting items. On the other, they enable different characteristics of the data to be taken into account, such as the order or time structure, which leads to significant improvements in performance.

Two recommendation systems were rolled out: one focussing on users, and the other concentrating on items. With the user-based focus, the users are recommended items that similar users like. With an items-based focus, the user receives recommendations for items that are similar to items they have liked in the past. Sciling’s ingenious idea was to build a hybrid system combining the two focuses to achieve the best possible result. “The users were amazed at the quality of the recommendations,” Otaku’s creator commented.

Technologies used

Recommendation and Predictive Analysis Systems

To associate images with users

Neural Networks

To predict users’ ratings


To train and implement the neural networks

Microservices with Docker

To create end-to-end solutions

Why us?

  • We have wide experience in developing recommendation systems.
  • We have been helping companies for years in all sectors to ensure visibility for their business model.
  • We can count on a team of researchers used to taking technology beyond the existing limits.
  • We have over 15 years’ experience in carrying out projects involving research into machine learning and its applications.