Project Overview
Our client operates a popular weight-loss and wellness app built around behaviour change — turning healthy habits into something entertaining, social and supported by expert voices. As their library of live sessions, audio clips, podcasts and trending topics grew, they needed a recommendation engine capable of matching the right content to each user's goals, mood and context in real time.
The Challenge
- Handle a heterogeneous content catalogue — live audio, on-demand podcasts, short clips, chats and text — each with very different metadata and signal patterns.
- Capture the subtle qualities that make audio valuable: not just the topic, but the answer itself, its relevance, the wisdom it carries, and the authority of the voice delivering it.
- Generate recommendations in real time, reacting to the user's current activity rather than relying only on yesterday's behavioural snapshot.
- Preserve user engagement and loyalty at scale, across millions of users with very different journeys and goals.
Our Approach
Multi-modal representation learning
We built embeddings for audio, text and structured session metadata, so heterogeneous content could be compared and ranked inside a shared space.
Siamese-inspired matching network
A siamese-architecture-inspired network considered multiple parameters — user goals, recent behaviour, content features and contextual signals — to score candidate items for relevance and likely engagement.
Real-time personalisation
Recommendations were recomputed continuously as the user interacted with the app, so the surfaced content reflected the current session rather than a stale batch score.
Feedback-aware ranking
Implicit signals — listens, drop-offs, replays, saves, shares — fed back into the ranking model so the system kept improving as the catalogue and audience evolved.
Technology Stack
The solution was engineered with a carefully chosen set of tools and frameworks, balancing maturity, performance and fit to the problem domain.
Results & Impact
+45% improvement in user retention
as users kept finding content that genuinely resonated with where they were in their journey.
Increase in daily active usage
with recommendations becoming a core reason users returned to the app.
Better content discovery
across audio, chats, podcasts and trending topics — content types users previously had trouble navigating on their own.
Positive impact at scale
with millions of users exposed to more relevant, personally meaningful guidance.
Conclusion
Recommendations in a wellness product are not just a growth lever — they are part of the product's duty of care. By investing in multi-modal representations, real-time personalisation and feedback-aware ranking, we helped the client build a system that consistently put the right voice, at the right moment, in front of the right user.