The model
Two models do the work: one understands your content the day it's created, and one learns from how people behave. Both retrain on your data, the learned one has to prove itself before it can take over, and you can put a new model up against the current one in a live test.
Two models, one engine
Learns from each item's own features and meaning — its type, language, categories, and what the text is actually about.
Learns a representation of each person and each item from real interactions, then ranks one against the other.
Training & quality
Training runs in the background — kick one off and it produces a new candidate model. Before that candidate ever serves a request, it's graded on held-out interactions the model never saw. A new model only takes over if it measures up to the one already serving; if it falls short, the current model keeps its place.
A/B testing
An offline grade is only half the story — what matters is whether people engage. Two ready-made experiments let you find out: put the model up against a simpler heuristic, or a new model version up against the current one. Each splits traffic by a feature flag and measures engagement and positive feedback.
Keep exploring
How candidates are generated, ranked live, and tuned by feedback.
The candidate generators — trending, co-engagement, and the live model.
Named surfaces that assemble a ranked set for any spot in your app.
Turn profile attributes and segments into what personalizes a feed.