Personalised Recommendations by Leveraging Generative AI with Calm
Enhance user experience and drive business value through increased retention, revenue, and lifetime value by leveraging Large Language Models to generate tailored recommendations and justifications.
Personalised Recommendations by Leveraging Generative AI with Calm
Enhance user experience and drive business value through increased retention, revenue, and lifetime value by leveraging Large Language Models to generate tailored recommendations and justifications.
At a glance
Calm is a software company based in San Francisco that produces a leading meditation and sleep application, helping users improve their overall mental health and wellbeing.
Challenge
Calm was looking to enhance its customer experience through improving its personalised content recommendations.
Solution
Combining a vector database, LLMs, and a hybrid search approach to generate personalised recommendations.
Services Used
- Amazon OpenSearch
- Amazon Bedrock
- AWS Lambda
- Amazon DynamoDB
Outcomes
- 100% of recommendations have explainability
- 25% more accurate recommendations
- 5 additional metadata properties for customisation
Business challenges
Improving personalisation of content recommendations
Calm recognised an opportunity to improve their personalised recommendations to customers and sought to evaluate the use of Large Language Models (LLMs) to enhance the personalisation of content and improve the user experience.
While Amazon Personalise was being used to assist with personalised recommendations, Calm wanted to increase performance and incorporate a layer of contextual awareness to the recommendation process. The goal was to leverage user data and content metadata to generate more relevant and engaging recommendations for their customers.
“I think the work that Firemind has done has been brilliant and it just continued to improve with each iteration. I love that even just at first glance, you could tell there was an improvement in quality.”
Jonathan Hummel, VP of Engineering — Calm
Solution
Integrating LLMs for personalised recommendations
The Firemind team implemented a solution that leveraged a combination of a vector database, LLMs, and a hybrid search approach to generate personalised recommendations for Calm users. This solution was designed to enhance the personalisation of content recommendations, ultimately driving improved business outcomes for Calm.
A key aspect of the solution was the generation of detailed content descriptions using LLMs. By incorporating the transcript data and metadata associated with Calm‘s content catalogue, the team was able to create more nuanced and contextual descriptions of each item. This provided the LLMs with richer information to draw upon when matching content to user preferences and personas.
The final step of the solution involved the LLMs providing justifications for the recommended content. These explanations, addressed directly to the user, highlighted how the suggested items complemented their mood and persona. This added layer of transparency and personalisation was designed to enhance the user experience and drive greater engagement with the recommended content.
Personalised user profile
The solution utilised the user's recent engagement with Calm's content to tailor recommendations for individuals based on their preferences.
Generated detailed content description
By incorporating the transcript data and metadata associated with Calm's content catalogue, the team was able to create more nuanced and contextual descriptions of each item, providing the LLMs with richer information to draw upon when matching content to user profiles.
Model Spotlight
Claude 3 Sonnet
We chose Claude 3 Sonnet for its superior ability to process and summarise large volumes of data, particularly transcribed audio from the Calm app.
Claude 3 Sonnet's large content window (200K) enabled it to handle extensive transcripts and metadata, generating detailed and nuanced user profiles. These profiles, which included user mood and persona, were crucial for the advanced RAG (Retrieval-Augmented Generation) system that powered Calm's recommendation engine.
By combining content summarisations with user profiles, Claude 3 Sonnet played a central role in recommending personalised content that aligned with users' interests and emotional states. Its ability to manage and distill large amounts of data efficiently made it the preferred model for comprehensive summarisation, especially when paired with Mistral for a more balanced approach.
This resulted in 100% of the recommendations being explainable, with clear justifications provided for each suggestion, improving transparency and user trust.
Mistral Large
We chose Mistral Large for its strong performance in handling smaller, more focused summarisations, particularly where speed and concise processing were needed.
Mistral was employed in areas where shorter content and metadata summaries were sufficient, effectively complementing Claude 3 Sonnet’s more expansive capabilities. By working together, Claude managed the larger, more detailed summarisations, while Mistral handled shorter, quicker summaries.
This combination ensured efficient performance across different data types, allowing us to use each model where it excelled, ultimately enhancing the advanced RAG framework.
The dual-model approach improved the accuracy of recommendations by 25%, as the system could better match user profiles with content based on both detailed and concise data summaries. Additionally, the implementation allowed for the inclusion of 5 additional metadata properties for further customisation, enabling a more tailored user experience.
Discover the Calm app
You can download the Calm app for both Google Play and the App store.
100%
of recommendations now come with explainability
The LLMs provided justifications for the recommended content, highlighting how the suggested items complemented the user's mood and persona.
25%
more accuracy for recommendations
The combination of personalised user profiles, detailed content descriptions, and the hybrid search approach led to a 25% improvement in the accuracy of the recommendations.
Added value
The solution added 5 new metadata properties per recommendation, such as title, description, mood-based, sleep-based, and duration, enabling more customisations and personalisation.
Get in touch
Want to learn more?
Seen a specific case study or insight and want to learn more? Or thinking about your next project? Drop us a message below!