Workflow · Production

Google builds a personalized AI health coach powered by Gemini models for Fitbit Premium users

The problem

Health and fitness journeys have historically been fragmented, generic, and inaccessible — providers suggest actions like seeing a specialist or losing weight without connecting users to the right resources, leaving users to connect the dots themselves.

Workflow diagram · grounded in source
1
User health question
trigger
“"Do I get better sleep after exercising?" sounds like a simple question, but answering it like a proactive, personalized and adaptive coach required several technical innovations”
2
Physiological data analysis
ai_action
“the coach verifies recent data availability, chooses the right metrics, contrasts relevant days, contextualizes results against personal baselines and population-level statistics, incorporates prior interactions with the coach, and final…”
3
Multi-agent coordination
ai_action
“we utilize a multi-agent framework that coordinates expert sub-agents to provide clear, consistent and holistic support, such as (1) a conversational agent for multi-turn conversations, intent understanding, agent orchestration, context …”
4
SHARP validation and human review
validation
“we continuously validate the personal health coach using dimensions of safety, helpfulness, accuracy, relevance, and personalization, collectively known as the SHARP evaluation framework. This multi-level assessment involves over 1 milli…”
5
Personalized plan output
output
“a domain expert, such as a fitness expert that analyzes user data to generate personalized fitness plans and adapt them as progress and context change”
Reported outcome

Google launched a public preview of the AI-powered personal health coach for eligible US-based Fitbit Premium users, validated through the SHARP evaluation framework involving over 1 million human annotations and more than 100k hours of human evaluation.

Reported metrics
Human annotations used in evaluationover 1 million
Human evaluation hoursmore than 100k hours
Reported stack
Gemini modelsFitbitSHARP evaluation framework
Source
https://research.google/blog/how-we-are-building-the-personal-health-coach/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Google launched a public preview of the AI-powered personal health coach for eligible US-based Fitbit Premium users, validated through the SHARP evaluation framework involving over 1 million human annotations and more…

What tools did this team use?

Gemini models, Fitbit, SHARP evaluation framework.

What results were reported?

Human annotations used in evaluation: over 1 million; Human evaluation hours: more than 100k hours (source-reported, not independently verified).

How is this workflow AI workflow structured?

User health question → Physiological data analysis → Multi-agent coordination → SHARP validation and human review → Personalized plan output.