How a lightweight recommendation model reshaped onboarding for a content app

AI · 6 min read

The startup faced typical cold-start friction: new users were asked to follow topics, but many skipped, leaving feeds blank or generic. Designers proposed embedding a tiny on-device model that inferred preferences from a short swipe test and device context, rather than waiting for server-side signals. The hypothesis: short, contextual predictions would create immediate relevance and encourage more feed interactions.

Product, design, and machine learning collaborated to define constraints — model size under 2 MB, latency under 50 ms, and no raw behavioral data sent off device until consented. The team built an A/B test where the control was the classic topic picker and the variant used a three-card swipe test whose responses were fed to the on-device model. Metrics tracked included first-session time-on-feed, follow rate, and permission opt-in for cross-device personalization.

The on-device approach increased first-session time-on-feed by 34% and doubled the follow rate during onboarding. Importantly, opt-in to server-side personalization rose as users experienced a more relevant feed first, lowering long-term acquisition costs. The case underscores a practical lesson: lightweight AI that supports UX goals can be implemented within strict privacy and performance budgets to improve immediate product experience.