Almond

Inspiration

To live. What a wonderful thing, right? Draw a picture of happiness for yourself, the moment you'd feel the most blissful, complete, per se. You look to your right, and your eyes mist over your mom. Next to her, your silly sister is making a face at you. You hear your dad coming from afar, complaining about his retirement again. It only becomes funnier to hear those rants after some time; you let a gentle chuckle out as you breathe the tender breeze in.

It is a dream to ascend to that euphoria. An increasingly rare euphoria that is, alas, though ironically, in the 21st century, as we reach the pinnacle of technology, launch mission after mission, build modular reactors, and diagnose the most troublesome diseases easier than ever with AI.

Each year, over 23 million people die from chronic diseases. At least 40% of these premature deaths are from chronic diseases that could have been prevented or delayed as simply as by early diagnosis. At least 9 million families that could have been in the picture you drew earlier. Meanwhile, low-income Americans are 2.4x more likely to skip the doctor entirely in a given year. We believe that the chance to be in that picture shouldn't depend on one's socio-economic status, so we built Almond.

What it does

Almond combines 30 seconds of onboarding questionnaire with data that your Smart Wearable Device (Apple Watch/Whoop Band/Oura Ring/or even the ITouch Smart Watch from Target for $50) already collects such as step count, sleep metrics, heart rate, etc. to give you a single Vitality Score that shows how your physiological health is doing as of today and where it's trending.

Behind our Vitality Score, we layered decades of US population health research with the same risk equations that specialists use to detect heart disease and diabetes, enabling you and your clinician to see your physiological health on a continuum.

Essentially, Almond is a preventative cardiovascular health platform that transforms everyday wearable data into personalized long term health insights.

How we built it

We trained a survival model on 8,500 US adults whose long-term outcomes are tracked in public NHANES data, layered 4 clinically-validated heart and diabetes risk equations (ASCVD, Framingham, FINDRISC, LE8) on top, and used Google's Gemma to write a personalized narrative for every score. The stack is a FastAPI + MongoDB backend that retains every snapshot append-only, a SwiftUI iOS app that streams Apple Watch data, and a Next.js dashboard built for clinicians.

Challenges we ran into

One challenge was that the NHANES public mortality data was only tracking subjects for 24 months, so we calibrated the survival model to its true support window and ship every prediction with the exact horizon the data can defend, essentially turning a public-dataset limitation into a *hard forecasting guarantee. *

NHANES also didn't contain heart-rate, HRV, or VO2 measurements, so we engineered an augmentation layer that fuses the trained Cox model with the largest published clinical studies for each signal (Jensen 2013, Hillebrand 2013, the FRIEND registry), giving every contribution a literature citation and a bounded effect on the Vitality Score

Accomplishments that we're proud of

We are proud that we trained a survival model on 8,500 real NHANES subjects with linked mortality outcomes, hit a C-index of 0.825 with clean calibration, and layered four clinically-validated risk equations (ASCVD, Framingham, FINDRISC, LE8) on top, every score on the dashboard is either learned from population-scale data or anchored to a peer-reviewed clinical equation.

We also shipped the full vertical end-to-end in one weekend:

A SwiftUI iOS app that streams Apple Watch HealthKit data, a FastAPI + MongoDB backend with append-only audit trails so every prediction is traceable to its source payload, and a typographically-distinct Next.js clinician dashboard whose design system is mirrored across iOS and web.

What we learned

We learned that the real challenge in AI-powered healthcare is not simply building a more complex model, but building a system people can genuinely trust. In preventative health, every prediction carries emotional and clinical weight, which means every score has to be explainable, defensible, and grounded in validated research rather than black-box AI claims.

We also learned how fragmented wearable health data still is. Millions of people already generate incredibly valuable cardiovascular signals every day through devices like the Apple Watch, yet most users — and even many clinicians — struggle to translate that data into meaningful long-term health understanding. Building Almond taught us that the true value is not just collecting health data, but transforming it into clear, actionable, and human-centered insights.

On the technical side, we learned that maintaining consistency across an end-to-end healthcare stack is as challenging as the machine learning itself. Synchronizing schemas and data contracts across Swift, Python, and TypeScript while preserving traceability and explainability became one of the most important parts of building a trustworthy platform.

What's next for Almond

The next architectural milestone is moving from literature-anchored wearable adjustments toward models trained directly on large-scale wearable datasets. With access to cohorts such as UK Biobank’s 500,000-subject wearable dataset with long-term follow-up, Almond will be able to learn cardiovascular relationships directly from heart rate, HRV, sleep, activity, and VO2 trends rather than relying on published augmentation studies alone.

Beyond the modeling layer, we plan to expand Almond into a clinician-connected preventative health platform. This includes multi-patient clinician dashboards, secure physician authentication, and FHIR-based integration with existing electronic health record systems to make wearable-derived insights easier to incorporate into real healthcare workflows.

Long term, we envision Almond becoming a passive early-awareness system capable of recognizing meaningful physiological deviations before symptoms become noticeable. By learning each user’s personal baseline over time, Almond could help identify unusual cardiovascular or recovery patterns days earlier through longitudinal wearable analysis and personalized anomaly detection.

Built With

• fastapi
• gemma
• mongodb
• next.js
• python
• swiftui
• typescript