This put up is co-written with Sunaina Kavi, AI/ML Product Supervisor at Omada Well being.
Omada Well being, a longtime innovator in digital healthcare supply, launched a brand new diet expertise in 2025, that includes OmadaSpark, an AI agent skilled with sturdy medical enter that delivers real-time motivational interviewing and diet training. It was constructed on AWS. OmadaSpark was designed to assist members establish their very own motivational challenges like emotional consuming, enhance meals choices, set targets, and maintain lasting conduct change. The next screenshot reveals an instance of OmadaSpark’s Dietary Training characteristic, demonstrating how members obtain customized diet training in actual time.
On this put up, we study how Omada partnered with AWS and Meta to develop this healthcare-aligned AI answer utilizing Llama fashions on Amazon SageMaker AI. We discover the technical implementation, structure, and analysis course of that helped Omada scale customized diet steerage whereas sustaining their dedication to evidence-based care.
The chance for AI-powered diet steerage
Vitamin training serves as a cornerstone of Omada’s persistent situation administration packages. Though well being coaches excel at offering customized care, the rising demand for fast, handy dietary info introduced a chance to boost our coaches’ affect by means of expertise. Omada sought an progressive answer that will complement their coaches’ experience by dealing with routine analytical duties, so they might focus extra deeply on significant member interactions. The aim was to offer instant, high-quality diet training whereas sustaining strict healthcare compliance with Omada’s care protocols and the private touches that makes their program efficient.
Omada Well being’s OmadaSpark goals to assist members establish real-world emotional and sensible obstacles to wholesome consuming in at this time’s setting, the place ultra-processed meals are prevalent and diets can fail to ship long-term outcomes. OmadaSpark options motivational interviewing,utilizing questions to assist members establish their very own targets, reinforce autonomy, and discover motivation to alter habits. OmadaSpark’s Dietary Training characteristic can cut back the psychological load of real-time meals choices and encourage members to step by step incorporate more healthy meals options. Omada’s diet expertise affords up to date monitoring capabilities, like water monitoring, barcode scanning, and photo-recognition expertise that supply versatile and non-restrictive help designed to advertise a wholesome relationship to meals.
“We see AI as a power multiplier for our well being coaches, not a substitute,” explains Terry Miller, Omada’s Vice President, Machine Studying, AI and Information Technique. “Our collaboration with AWS and Meta allowed us to implement an AI answer that aligns with our values of evidence-based, customized care.”
Answer overview
Omada Well being developed the Dietary Training characteristic utilizing a fine-tuned Llama 3.1 mannequin on SageMaker AI. The implementation included the Llama 3.1 8B mannequin fine-tuned utilizing Quantized Low Rank Adaptation (QLoRA) strategies, a fine-tuning methodology that enables language fashions to effectively study on smaller datasets. Preliminary coaching used 1,000 question-answer pairs created from Omada’s inner care protocols and peer reviewed literature and specialty society tips to offer evidence-based dietary training.
The next diagram illustrates the high-level structure of Omada Well being’s Llama implementation on AWS.
The answer workflow consists of the next high-level steps:
- The Q&A pairs for dietary training datasets are uploaded to Amazon Easy Storage Service (Amazon S3) for mannequin coaching.
- Amazon SageMaker Studio is used to launch a coaching job utilizing Hugging Face estimators for fine-tuning Llama 3.1 8B mannequin. QLoRA strategies are used to coach the mannequin and mannequin artifacts saved to Amazon S3.
- The inference workflow is invoked by means of a person query by means of a cell consumer for OmadaSpark’s dietary training characteristic. A request is invoked to fetch member private knowledge based mostly on the person profile in addition to dialog historical past, in order that responsive info is customized. For instance, a roast beef recipe received’t be delivered to a vegetarian. On the similar time, this characteristic doesn’t present medical info that’s associated to a specific particular person’s medical scenario, corresponding to their newest blood glucose take a look at. The SageMaker AI endpoint is invoked for diet era based mostly on the member’s question and historic conversations as context.
- The mannequin generates customized diet training, that are fed again to the cell consumer, offering evidence-based training for individuals in Omada’s cardiometabolic packages..
- For analysis of the mannequin efficiency, LangSmith, an observability and analysis service the place groups can monitor AI utility efficiency, is used to seize inference high quality and dialog analytics for steady mannequin enchancment.
- Registered Dietitians conduct human assessment processes, verifying medical accuracy and security of the diet training offered to customers. Upvoted and downvoted responses are seen in LangSmith annotation queues to find out future fine-tuning and system immediate updates.
The next diagram illustrates the workflow sequence in additional element.
Collaboration and knowledge fine-tuning
A essential facet of Omada Well being’s success with AI implementation was the shut collaboration between their medical group and the AI growth group. Omada AI/ML Product Supervisor Sunaina Kavi, a key determine on this collaboration, highlights the significance of this synergy:
“Our work with the medical group was pivotal in constructing belief and ensuring the mannequin was optimized to fulfill real-world healthcare wants,” says Kavi. “By intently engaged on knowledge choice and analysis, we made certain that OmadaSpark Dietary Training not solely delivered correct and customized diet e but additionally upheld excessive requirements of affected person care.
“The AWS and Meta partnership gave us entry to state-of-the-art basis fashions whereas sustaining the self-hosted management we want in healthcare, for privateness, safety, and high quality functions. The fine-tuning capabilities of SageMaker AI allowed us to adapt Llama to our particular diet use case whereas preserving our knowledge sovereignty.”
Affected person knowledge safety remained paramount all through growth. Mannequin coaching and inference occurred inside HIPAA-compliant AWS environments (AWS is Omada’s HIPAA Enterprise Affiliate), with fine-tuned mannequin weights remaining below Omada’s management by means of mannequin sovereignty capabilities in SageMaker AI. The AWS safety infrastructure offered the inspiration for implementation, serving to keep affected person knowledge safety all through the AI growth lifecycle. Llama fashions provided the flexibleness wanted for healthcare-specific customization with out compromising efficiency. Omada centered their technical implementation round SageMaker AI for mannequin coaching, fine-tuning, and deployment.
Lastly, Omada applied rigorous testing protocols, together with common human assessment of mannequin outputs by certified. Omada launched the complete workflow with the mannequin in 4.5 months. All through this course of, they repeatedly monitored response accuracy and member satisfaction, with iterative fine-tuning based mostly on real-world suggestions.
Enterprise affect
The introduction of OmadaSpark considerably boosted member engagement of those who used the software. Members who interacted with the diet assistant have been thrice extra more likely to return to the Omada app normally in comparison with those that didn’t work together with the software. By offering round the clock entry to customized dietary training, Omada dramatically lowered the time it took to deal with member diet questions from days to seconds.
Following their profitable launch, Omada is deepening their partnership with AWS and Meta to increase AI capabilities together with fine-tuning fashions, context window optimization, and including reminiscence. They’re growing a steady coaching pipeline incorporating actual member questions and enhancing AI options with further well being domains past diet.
“Our collaboration with AWS and Meta has proven the worth of strategic partnerships in healthcare innovation,” shares Miller. “As we glance to the long run, we’re excited to construct on this basis to develop much more progressive methods to help our members.”
Conclusion
Omada Well being’s implementation demonstrates how healthcare organizations can successfully undertake AI whereas addressing industry-specific necessities and member wants. Through the use of Llama fashions on SageMaker AI, Omada amplifies the humanity of well being coaches and additional enriches the member expertise. The Omada, AWS, and Meta collaboration showcases how organizations in extremely regulated industries can quickly construct AI purposes through the use of progressive basis fashions on AWS, the trusted healthcare cloud supplier. By combining medical experience with superior AI fashions and safe infrastructure, they’ve created an answer that may rework care supply at scale whereas sustaining the customized, human-led method that makes Omada efficient.
“This venture proves that accountable AI adoption in healthcare isn’t just potential—it’s important for reaching extra sufferers with high-quality care,” concludes Miller.
Omada stays dedicated to rising its human care groups with the effectivity of AI-enabled expertise. Wanting forward, the group is devoted to creating new improvements that foster a way of real-time help, confidence, and autonomy amongst members.
For extra info, see the next sources:
In regards to the authors
Sunaina Kavi is an AI/ML product supervisor at Omada, devoted to leveraging synthetic intelligence for conduct change to enhance outcomes in diabetes, hypertension, and weight administration. She earned a Bachelor of Science in Biomedical Engineering and an MBA from the College of Michigan’s Ross Faculty of Enterprise, specializing in Entrepreneurship and Finance. Previous to transitioning to Omada, she gained expertise as an funding banker in Expertise, Media, and Telecom in San Francisco. She later joined Rivian, specializing in charging options inside their infotainment group, and based her personal startup aimed toward utilizing AI to handle autoimmune flares. Sunaina can be actively concerned within the Generative AI group in San Francisco, working to boost security, safety, and systematic evaluations inside the healthcare neighborhood.
Breanne Warner is an Enterprise Options Architect at Amazon Internet Companies supporting healthcare and life science (HCLS) prospects. She is enthusiastic about supporting prospects to make use of generative AI on AWS and evangelizing mannequin adoption for first-party and third-party fashions. Breanne can be Vice President of the Ladies at Amazon with the aim of fostering inclusive and numerous tradition at Amazon. Breanne holds a Bachelor of Science in Pc Engineering from the College of Illinois Urbana-Champaign.
Baladithya Balamurugan is a Options Architect at AWS targeted on ML deployments for inference and utilizing AWS Neuron to speed up coaching and inference. He works with prospects to allow and speed up their ML deployments on companies corresponding to Amazon SageMaker and Amazon EC2. Based mostly out of San Francisco, Baladithya enjoys tinkering, growing purposes and his homelab in his free time.
Amin Dashti, PhD, is a Senior Information Scientist at AWS, specializing in mannequin customization and coaching utilizing Amazon SageMaker. With a PhD in Physics, he brings a deep scientific rigor to his work in machine studying and utilized AI. His multidisciplinary background—spanning academia, finance, and tech—permits him to deal with complicated challenges from each theoretical and sensible views. Based mostly within the San Francisco Bay Space, Amin enjoys spending his free time along with his household exploring parks, seashores, and native trails.
Marco Punio is a Sr. Specialist Options Architect targeted on GPU-accelerated AI workloads, large-scale mannequin coaching, and utilized AI options on AWS. As a member of the Gen AI Utilized Sciences SA group at AWS, he makes a speciality of high-performance computing for AI, optimizing GPU clusters for basis mannequin coaching and inference, and serves as a world lead for the Meta–AWS Partnership and technical technique. Based mostly in Seattle, Washington, Marco enjoys writing, studying, exercising, and constructing GPU-optimized AI purposes in his free time.
Evan Grenda Sr. GenAI Specialist at AWS, the place he works with top-tier third-party basis mannequin and agentic frameworks suppliers to develop and execute joint go-to-market methods, enabling prospects to successfully deploy and scale options to unravel enterprise agentic AI challenges. Evan holds a BA in Enterprise Administration from the College of South Carolina, a MBA from Auburn College, and an MS in Information Science from St. Joseph’s College.
