The media world is changing in leaps and bounds. More and more listeners prefer to listen to content by demand and adapted to suit their tastes than use traditional systems. When listening to podcasts, it is hard for them to find new content that they might like. This is why a recommender is so important.
How can we give each listener a personalised response? Tailorcast, a podcast recommender that is capable of suggesting new content based on each user’s tastes or other content by taking the characteristics and content of each episode and the settings of each recommendation channel into account, is born out of this context.
PRISA presented this conceptualised idea to us, and we at Paradigma have travelled with them ever since, advising them on technical decisions and adapting to the changes in the project, working with them as a single team.
A connected, multi-purpose recommender
The Tailorcast solution consists of two main parts: an audio preprocessing motor, where the audios are converted into text and sorted into categories, and an API that permits getting recommendations from any front-end or system in the form of episodes and lists of personalised podcasts.
Tailorcast is connected to the Group’s content repository. Thus, the meta-information of the audios is enriched and stored in a single location, thus facilitating synergies and the use of the recommender among the company’s different departments.
We needed a marketplace system that allowed content generators to upload their content to the platform and configure it. It was also necessary for publishers to be able to configure their channels, dynamically indicating which podcasts could be added to playlists.
Another thing that had to be taken into account was the need for an SDK that allowed the mobile applications of publishers to link to the Tailorcast recommendation system.
Personalised lists of episodes and podcasts thanks to AI
Main parts of the architecture
After defining the scope of the MVP, we began to work on assembling the different parts of the architecture:
A backoffice API microservice with which the API provides service to the front-end.
An SDK-API microservice for taking calls not only from mobile SDKs but from any front-end.
SDK libraries that can be integrated with iOS and Android apps so as to make use of the recommender.
Demo apps with a built-in library.
A podcast recommendation motor based on content and the listening data for clustering users and generating recommendations.
In order to be able to make recommendations, we created an audio processing pipeline that validates, analyses and extracts information from the audios thanks to GCP’s AI APIs.
Google Cloud, a travel companion
We developed this project in Google Cloud using managed serverless services such as App Engine, Cloud Run, PubSub, BigQuery, Cloud SQL and AI training. In addition, we used ML APIs to transcribe the audios, extract entities and get the sentiment. We also trained an entity extraction model using the Group’s labelling system via the AutoML service.
The microservices were developed using Node.js, adding Swagger, and the front-end with React. We used IaC in order to be able to easily deploy it with Terraform.
A better experience for the user
Tailorcast has multiple benefits. With the independent preprocessing and recommendation motor, PRISA’s companies can load and process their audios and get content recommendations for their users.
The preprocessing provides the transcription, sentiment analysis, classification and validation of the audios fed into the system. By saving the information in an analytical database, the systems can utilize this meta-information separately from the recommendation proper.
Additionally, it allows each user to create their own personalised playlists and, thanks to its multichannel approach, it mixes content from different creators to provide service to customers from different media.
Thanks to this intelligent, personalised recommendation, users will have content that is closer to their tastes and enjoy an enhanced experience by discovering new podcasts.
In addition to enthusiasm and commitment to the project, Paradigma provided methodological consistency and its expert knowledge of the Google Cloud environment.