try it

logged out, you can still import an Apple Health export: it's parsed in your browser and nothing leaves it. the full flow needs an account and the iPhone app.

overview

HuskyHealth is what I built solo over a weekend at HACK1984, Northeastern's hackathon run by the Claude Builders Club, Rev, ACM and AINU on November 7–8, 2026. It won 3rd place in the Privacy-First track, the main track ("software that protects its users by design"), with a $1,000 cash prize. The full submission is on Devpost, and the demo video is below.

The tagline is "zero knowledge, personal API." Your watch knows how hard you trained, how you slept and how fit you are, but getting that data out means either exporting a zip by hand or giving your health history to a company whose server reads it. HuskyHealth sends every Apple Watch or Garmin workout to your own private REST API automatically, and nobody can read it on the way, not even the server that stores it. On top of that API sit a daily training call from published sports science, a website and an iPhone app that show the same numbers, and a voice you can talk to. As a club squash player juggling training and classes, I built it for athletes in season first.

features

watch to API, by itself: Finish a workout and the iPhone app picks it up through HealthKit background delivery, seals it on the phone and posts it to your API. Garmin works too, through Garmin Connect's Apple Health sync.
zero knowledge: Your password never leaves your device. The server stores only ciphertext and a hash; the keys to open your history exist only on your phone and in your browser.
today's call: "Train hard" or "go easy", from your heart-rate variability band, or from this week's load against your usual week when there isn't enough HRV yet. Every number links to the paper behind it.
the same on every screen: The website and the app draw from the same encrypted history with the same code: this week, last session, sleep, VO2 max, heart-rate zones and calories.
talk to your API: Tap listen and today's call is read aloud with each word highlighted as it's said, using Gradium's text-to-speech. Ask "how many calories did I burn this week?" out loud and it's answered from your data, with the exact curl request behind the answer shown underneath.
no location, ever: Routes, ECGs and clinical records are never read.

the app

The iPhone app and the website share one design: a dot-matrix printout in Space Mono, with ./today, ./api, ./methods and ./privacy as the tabs. Today's call sits at the top with listen and speak under it, then a row for each number, and every row opens the same detail page on both. The privacy page spells out who can read what.

The HuskyHealth iPhone app's today screen: TRAIN HARD in dot-matrix letters, the reason (this week's load is 80% of your usual week), listen and speak, the ask prompt, the cited method and heart-rate zones
The last session page: a 101-minute squash session with its load, calories, comparison to the usual squash session and minutes in each heart-rate zone The privacy page: a table of who can read the password, the username, the workouts and the key that opens them (only you), and location, routes and ECGs (never read)

how the privacy works

accounts: The browser or phone runs PBKDF2-SHA256 (600,000 rounds) on your password. Half of the result is the vault key, which never leaves your device. The other half logs you in, and the server keeps only a scrypt hash of it.
the vault: Your history is JSON, gzipped, then encrypted with AES-256-GCM before it leaves the device.
write-only device tokens: Records from devices go in with a token that can add but never read. Each one is sealed to your account's RSA public key (AES-256-GCM, with the key wrapped by RSA-OAEP); the private key lives only inside your encrypted vault.
see what leaves: Voice requests go out through an egress proxy, so you can watch exactly what reaches Gradium: one short passage of the brief, or at most five seconds of your question, never stored.
the honest cost: There's no password reset. Nobody can read your history, so nobody can recover it for you either.

the engine

The training call uses published methods only: Banister TRIMP and the fitness-fatigue model, session-RPE, HRV bands from Plews et al., and the acute:chronic load ratio from Gabbett, labelled as disputed wherever it's used. The engine is a single JavaScript file that runs in the browser and, as the same file, in the app, which is why both always show the same numbers. It's training insight, not medical advice.

construction

JavaScript Web Crypto Web Workers PWA Expo (React Native) TypeScript Swift HealthKit Python Flask nginx mitmproxy Docker Raspberry Pi 5 Cloudflare Pages Gradium node:test pytest

The website is a plain-JavaScript PWA with no framework: an Apple Health export is parsed in a Web Worker, and all the cryptography is the browser's own Web Crypto. The iPhone app is Expo with a Swift module for HealthKit, the Keychain, crypto and voice, and it reuses the website's JavaScript for the engine and the views. The API is a Flask app in Docker behind an nginx gateway with a method allowlist, per-client rate limits and a strict schema, on the same Raspberry Pi 5 as my homelab. 194 automated tests cover the engine, the parser, the vault, the views and the API, including a check that nothing gets logged.

architecture

Apple Watch / Garmin → Apple Health → iPhone app (seals each workout) → Flask API on the Pi (ciphertext only) → website + app (decrypt on your device)

The server in the middle never holds anything it can read. Everything that turns ciphertext back into workouts happens at the ends: on your phone, in your browser, or on your own machine with the small command-line tool that reads your whole history back.

Source? It's on GitHub. The demo is on YouTube and the write-up is on Devpost.