Modulai teamed up with Frever to create an end-to-end machine-learning-based feed recommender system for their video-creation and social content sharing app. The multi-model architecture populates the feeds of every user, taking both video content and user preferences into account to ensure the best possible experience and relevance.
Challenge
In Frever's app, users create and share short animated videos by building avatars and scenes, adding voice and music. The videos are shared and discovered through a TikTok-like feed where users can share, comment, like, and remix. To offer the best possible user experience, a relevant and personalized feed that captures engagement and stimulates creativity was essential.
Solution
Engineers at Frever and Modulai built the recommender as a multi-model system, combining content-based models that read what each video actually contains with collaborative filtering that learns from user behavior. Together they score and rank videos for each user's feed, balancing what a user has shown interest in against the content of videos they have not seen yet. The aim was a feed that stays personal and relevant while still surfacing new content worth engaging with.
Tools
The solution was fully cloud-based, covering training, validation, and deployment. Development was mainly done in Python as dockerized microservices, with a set of internal APIs passing input data and recommendations between the producing and consuming systems.
Value created
The recommender powers the feed every Frever user sees, ranking videos in real time from both their behavior and the content of the videos themselves. For an app built around short creative videos, that means a feed tuned to keep users engaged and to give newly created content a chance to be discovered, rather than one driven by popularity alone.
Learn more
