How Vitality AI Health Actually Works: The Architecture Behind Personalized Health Intelligence

How Vitality AI Health Actually Works: The Architecture Behind Personalized Health Intelligence

Here is a question that took me about two years to figure out how to answer: why does your health data feel so useless?

Not because the data is wrong. Your Oura Ring Gen 4 is measuring your HRV down to the millisecond. Your Garmin Fenix 8 is logging your VO2 max every run. Your last blood panel probably has 40 markers on it. You have data. You just do not have a way to think across it. Each source gives you one reading on one dimension, and none of them talk to each other.

That gap is what I built Vitality AI Health to close. I am Kevin Amrelle, the founder. This post walks through the actual architecture of the platform: what data we connect, how the system weights it, and what the experience of using it actually looks like once the pieces are in place.

The Data Layer: What We Actually Connect

The premise of the platform is simple to state and operationally complicated to execute: pull every meaningful data source a person has about their health into one place, normalize it, and build a shared intelligence layer on top of it.

On the wearable side, we currently support Oura Ring Gen 4, Whoop 5.0, Garmin Fenix 8, Apple Watch Series 11, and Fitbit Sense 3, plus Apple Health as an aggregation layer for any device that routes through it. These devices are not interchangeable. Each one has a different strength. The Oura Ring Gen 4 leads on sleep staging accuracy and nocturnal HRV. Whoop 5.0 leads on strain quantification for training athletes. Garmin Fenix 8 gives the most granular GPS and performance data for endurance sports. Apple Watch Series 11 is the broadest-coverage consumer device, particularly for ECG and blood oxygen. When you connect more than one, the system cross-validates readings where the signals overlap, which turns out to matter: a 2025 review in npj Digital Medicine found that integrating data from multiple wearable streams allows AI algorithms to provide guidance that single-device approaches cannot reach on their own (Mahajan, Heydari, Powell. npj Digital Medicine, 2025. DOI: 10.1038/s41746-025-01554-w).

On the lab and clinical side, users can upload blood panel PDFs, standard results from any lab, and the system parses the markers automatically. The same upload flow works for DEXA scan results (body composition: lean mass, fat mass, bone density, visceral fat estimate) and DNA results exported from 23andMe in standard format. When you layer those three data types together, you get something that no wearable can approximate: a structural picture of your body that changes slowly over time, sitting underneath the daily signal your devices generate.

We also connect MyFitnessPal for nutrition logging. Dietary data is the weakest link in most health tracking, it depends entirely on user consistency, but when it is there, the system uses it to contextualize recovery patterns and energy availability.

The VAI Score: Why Data Completeness Is the Whole Game

The platform generates a composite number called the VAI Score. It reflects how well the nine data categories are populated and what they collectively show about your health status. Those nine categories are: Wearables, Blood Tests, DEXA Scans, DNA Results, Nutrition, Supplements, Medications, Environment, and Health Profile.

The reason we structured it across nine categories is not arbitrary. I spent a long time mapping every type of consumer-accessible health data and collapsing the redundancies. Blood and DEXA are intentionally separated because they measure different things: one is biochemical state, the other is structural composition. A person with excellent blood markers but undetected high visceral fat has a different risk profile than their blood work alone would suggest.

What the VAI Score actually reflects is data completeness multiplied by what that data shows. Connecting one device gives you a starting number. Uploading your last two blood panels pushes it higher. Adding a DEXA result and your 23andMe file pushes it higher still. This is not a gimmick. The quality of the AI's recommendations depends directly on the completeness of the input. A JAMA-recognized study analyzing 519 healthcare AI papers found that only about 5% used real patient data, most models were trained and tested on proxies that do not reflect how health AI actually performs in the real world. Vitality AI Health runs against your actual data, not a simulated profile.

I want to give you a concrete example of what completeness looks like in practice. When I first connected my own Garmin Fenix 8, my VAI Score came in at 61. Reasonable wearable data, nothing else. The week I uploaded my last two blood panels, it moved to 84. Not because the AI suddenly thought I was healthier, but because it now had enough context to answer specific questions: why my sleep recovery score was lagging on high-training weeks, what my ferritin trend suggested about my iron intake, whether my HRV trajectory over three months was consistent with the other markers. The score going up meant the intelligence got sharper, not that I passed some arbitrary threshold.

Vaya Chat: Where the Intelligence Surface Is

The part of the platform that most users interact with most is Vaya Chat, our AI health companion. Vaya Chat is the interface through which you ask questions, explore your data, and get responses that are grounded in what the platform actually knows about you rather than in generic health advice.

The distinction matters. A general-purpose AI can tell you that ferritin under 30 ng/mL is associated with fatigue. Vaya Chat can tell you that your ferritin has been declining across your last three panels, that your training load has been high over the same period, and that your sleep HRV has tracked in the same direction. That convergence is a specific finding, not a generic answer.

Vaya Chat also cites its sources. Every substantive claim references the study or guideline behind it, which is how we think health AI should work. The platform is not in the business of making claims that cannot be traced back to evidence. That is a design choice, not a legal disclaimer.

The Architecture Behind the Personalization

Understanding why the platform works requires understanding what "personalized" actually means technically, as opposed to how the word gets used in marketing copy.

Personalization in our context means: the model's outputs change based on your specific inputs, not based on demographic averages or population-level associations. When the system evaluates your recovery, it is comparing your last night's HRV against your own baseline, not against a population percentile. When it surfaces a pattern in your bloodwork, it is doing so relative to your own longitudinal trajectory, not against an age-matched reference range alone.

This architecture is why the choice of wearable affects what the platform can do. Whoop 5.0 generates training strain data with a methodology that differs meaningfully from Garmin's Training Readiness score. When both are connected, the system has two strain models to triangulate from. When only one is connected, it works with one. Neither is wrong. One is more complete.

The platform launched on BetaList in April 2026 and has collected 100+ early access requests. We are releasing mobile apps for iOS and Android in May 2026. In the meantime, the web platform is live, and the full data connection flow, wearables, lab uploads, DEXA, DNA, is available to new users now.

What This Costs and What You Get

The platform is $10 per month or $89 per year. With the code VITALITY20, the annual plan is $69. That covers the full data layer, all device connections, unlimited lab uploads, DEXA and DNA parsing, Vaya Chat, VAI Score, and the complete dashboard.

For context on where the market is heading: Rock Health's turn-of-2026 report tracked Oura's 2025 fundraise at nearly $11 billion, the largest digital health valuation on record since tracking began in 2011. Consumer appetite for continuous, personalized health data is not a trend; it is a structural shift in how people relate to their own health. The wearables now exist. The lab infrastructure now exists. The missing piece has been a platform that treats all of those inputs as parts of one coherent picture.

That is what we built.

One More Thing

We also fully funded a school through Pencils of Promise. This matters to us as a company because health and education are not separate problems, they both come down to access and long-term investment. We plan to keep funding programs like it as the company grows.

If you want to dig deeper on what the platform shows once it is fully populated, start with the biological age post, which covers the science behind the longitudinal markers the platform tracks. And if you have questions about data privacy, specifically whether your health data should ever go into a general-purpose AI, the ChatGPT and health data post covers that directly.

The architecture is only as good as the data you put into it. If you have a blood panel sitting in your email from your last annual physical, that is the fastest way to start.

Sources:

• Mahajan A, Heydari K, Powell D. Wearable AI to enhance patient safety and clinical decision-making. npj Digital Medicine. 2025;8(1):176. https://www.nature.com/articles/s41746-025-01554-w

• Rock Health. Healthcare innovation at the turn of 2026: Mapping what's now and what's next in digital health. December 2025. https://rockhealth.com/insights/healthcare-innovation-at-the-turn-of-2026-mapping-whats-now-and-whats-next-in-digital-health/

• Feldman MJ, et al. Dedicated AI Expert System vs Generative AI With Large Language Model for Clinical Diagnoses. JAMA Network Open. 2025;8(5):e2512994. https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2834550