AWS Architecture Blog·October 29, 2026Designing a Personalized AI Fitness Coach with Serverless and LLMs on AWS
This article details how boAt developed its "AI Coach" feature for smartwatches, leveraging AWS serverless technologies and Large Language Models (LLMs) to provide personalized fitness plans. It covers the architectural design choices focusing on personalization, affordability, and scalability, highlighting the use of Amazon Bedrock and AWS Lambda for real-time plan generation and adherence nudges. The architecture emphasizes a model-agnostic approach and zero infrastructure overhead, crucial for a high-scale consumer application.
Read original on AWS Architecture BlogboAt's AI Coach feature provides personalized fitness coaching plans to smartwatch users, utilizing user history, preferences, and generative AI. The core design principles for this system were personalization, affordability, and scalability, which guided the technology choices on AWS.
High-Level Architecture Overview
The AI Coach system collects data from boAt smartwatches and a companion mobile app. This data, including steps, activities, sleep, communication style, and fitness goals, is stored historically in Amazon RDS for MySQL and Amazon DocumentDB (with MongoDB compatibility). The core logic for plan generation and nudges is handled by AWS Lambda functions interacting with Amazon Bedrock.
Key Architectural Components and Their Roles
- Data Sources: boAt smartwatches capture metrics, while the companion app collects user preferences and goals. All data is encrypted.
- Historical Data Storage: Amazon RDS for MySQL and Amazon DocumentDB store comprehensive user history, providing crucial context for LLM personalization.
- AWS Lambda: Acts as a translation layer, transforming input data for LLMs and parsing LLM outputs for the app. It also handles business logic for plan generation and nudges. It abstracts the underlying AI coach implementation, allowing future changes to user inputs without backend modifications.
- Amazon Bedrock: Utilized with the Meta Llama 3.3 70B Instruct model for personalized plan generation. It offers model-agnostic access, built-in observability via CloudWatch, and serverless scalability with on-demand inference, charging per token processed. Plans for future transition to batch inference for cost optimization at scale.
Design for Flexibility
The architecture emphasizes abstracting the LLM implementation through a Lambda translation layer. This allows evaluating and switching between different foundation models without extensive backend changes, promoting future adaptability and reducing vendor lock-in risk for AI components.
Scalability and Cost-Effectiveness
The choice of AWS Lambda and Amazon Bedrock in on-demand inference mode ensures serverless scalability with zero infrastructure overhead. Bedrock scales automatically and charges per token, making it cost-effective for varying usage. Integration with Amazon CloudWatch provides precise per-invocation metrics, enabling detailed cost governance and unit economics tracking without custom metering infrastructure. Future plans include using cross-Region inference profiles for dynamic request routing to further enhance throughput and resilience.