By Teresa Torres · producttalk.org · @ttorres on X · LinkedIn
Teresa Torres interviews the founders of Rest, a startup that built a voice-first AI sleep coach grounded in Cognitive Behavioral Therapy for Insomnia (CBTI). The team began by noticing that users of their podcast app listened to content to fall asleep, which led them to a dedicated sleep audio product and, once LLMs emerged, to an AI coach. The conversation covers how they moved from a basic chatbot to a system with memory, dynamic daily agendas, and retrieval-augmented generation, while staying on the wellness side of the wellness-versus-medical line. Their incremental, error-driven approach to building a complex, personal AI product offers practical lessons for product teams. The source is a podcast episode page; the transcript is paywalled, so this summary draws on the show notes and description.
01Key takeaways
- Look for unexpected user behavior in existing products, such as off-label usage, as a signal for a new high-value segment.
- Ground an AI product in an established, evidence-based method rather than inventing guidance from scratch.
- Build AI capabilities one step at a time, shipping a simple version before adding voice, memory, or retrieval.
- Replace oversized system prompts with retrieval for general knowledge, keeping only user-specific data in the prompt.
- Define explicit safety boundaries for health-adjacent products and test them with automated evals.
- Run regular error analysis with domain experts to turn failures into concrete product iterations.
02Key sections
- Origin from user behavior
- The team spotted that a meaningful share of podcast app users were listening to fall asleep, a signal that pointed to a high-intent, paying use case. Jobs-to-be-done research then surfaced an underserved group of do-it-yourself sleep hackers.
- Choosing a clinically grounded foundation
- Rather than invent a sleep approach, they adopted CBTI principles because the method is clinically validated and effective. This gave the product a credible, evidence-based core.
- Evolving from text chatbot to voice
- The product moved from a text chatbot to a voice-first experience, with parallel development paths for text and voice. Early prototypes and error analysis drove iterative improvements.
- Memory, agendas, and retrieval
- Memory stores user context with time-based relevance, dynamic agendas shape each day's conversation from sleep data and program stage, and RAG replaced massive system prompts for general sleep knowledge.
- Guardrails and evaluation
- The team stayed clear of diagnosis and medication advice, ran weekly error reviews with sleep therapists, and built LLM-powered evals to test safety boundaries.
03From the post
“Listen to this episode: Spotify | Apple Podcasts What if you could get personalized sleep coaching—inspired by the same principles that cost thousands of dollars and have year-and-a-half waitlists—through a voice AI that checks in with you every morning? In this episode of Just Now Possible,”
04Frameworks mentioned
Summary and takeaways written by PM Atlas; quotes are short excerpts. © the original author.