Customer support · Production

Omada Health scales personalized nutrition care by fine-tuning Llama 3.1 on Amazon SageMaker AI

The problem

Health coaches at Omada excelled at personalized care, but growing member demand for quick nutritional information created an opportunity to supplement coach capacity with technology that could handle routine analytical tasks and provide immediate nutrition education at scale.

Workflow diagram · grounded in source

Training data uploaded to S3

integration

“The Q&A pairs for nutritional education datasets are uploaded to Amazon Simple Storage Service (Amazon S3) for model training.”

Llama 3.1 8B fine-tuned with QLoRA

ai_action

“The implementation included the Llama 3.1 8B model fine-tuned using Quantized Low Rank Adaptation (QLoRA) techniques, a fine-tuning method that allows language models to efficiently learn on smaller datasets. Initial training used 1,000 …”

Member question triggers inference

trigger

“The inference workflow is invoked through a user question through a mobile client for OmadaSpark's nutritional education feature.”

Fetch member profile and history

integration

“A request is invoked to fetch member personal data based on the user profile as well as conversation history, so that responsive information is personalized. For example, a roast beef recipe won't be delivered to a vegetarian. At the sam…”

SageMaker endpoint generates nutrition education

ai_action

“The SageMaker AI endpoint is invoked for nutrition generation based on the member's query and historical conversations as context.”

Personalized education delivered to app

output

“The model generates personalized nutrition education, which are fed back to the mobile client, providing evidence-based education for people in Omada's cardiometabolic programs.”

LangSmith monitors inference quality

validation

“LangSmith, an observability and evaluation service where teams can monitor AI application performance, is used to capture inference quality and conversation analytics for continuous model improvement.”

Registered Dietitians review outputs

human_review

“Registered Dietitians conduct human review processes, verifying clinical accuracy and safety of the nutrition education provided to users.”

Feedback drives future fine-tuning

feedback_loop

“Upvoted and downvoted responses are viewed in LangSmith annotation queues to determine future fine-tuning and system prompt updates.”

Reported outcome

OmadaSpark drove significantly higher member engagement — members who used the nutrition assistant were three times more likely to return to the Omada app — and reduced nutrition question response time from days to seconds, with the entire workflow launched in 4.5 months.

Reported metrics

Member app return ratethree times more likely to return to the Omada app

Nutrition question response timefrom days to seconds

Workflow launch time4.5 months

Initial training dataset size1,000 question-answer pairs

Reported stack

Llama 3.1 8BQLoRALangSmithHugging Face estimatorsMeta

Source

https://aws.amazon.com/blogs/machine-learning/how-omada-health-scaled-patient-care-by-fine-tuning-llama-models-on-amazon-sagemaker-ai?tag=soumet-20

Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

OmadaSpark drove significantly higher member engagement — members who used the nutrition assistant were three times more likely to return to the Omada app — and reduced nutrition question response time from days to se…

What tools did this team use?

Llama 3.1 8B, QLoRA, LangSmith, Hugging Face estimators, Meta.

What results were reported?

Member app return rate: three times more likely to return to the Omada app; Nutrition question response time: from days to seconds; Workflow launch time: 4.5 months; Initial training dataset size: 1,000 question-answer pairs (source-reported, not independently verified).

How is this customer support AI workflow structured?

Training data uploaded to S3 → Llama 3.1 8B fine-tuned with QLoRA → Member question triggers inference → Fetch member profile and history → SageMaker endpoint generates nutrition education → Personalized education delivered to app → LangSmith monitors inference quality → Registered Dietitians review outputs → Feedback drives future fine-tuning.