Back to Work
Mobile App 2024 — 18 weeks

Helio — The AI-Powered Wellness Companion That Users Actually Stick With

We designed and engineered a cross-platform health tracking app from zero — wearable sync, an on-device AI coaching engine, and deeply personalised insights that turned passive data collection into real behaviour change.

100k+
Active users
4.8
App Store rating
68%
Day-30 retention
18wk
Delivery time
Good morning, Alex ☀️
82
Wellness Score
Steps ↑
Sleep ✓
AI Coach
Your HRV trend is up 12% this week. Great recovery! 🎯
How can I improve my sleep score?
Try dimming screens 45 min earlier. You fell asleep 22 min faster on nights you did.

Turning health data into habits that stick

The wellness app market is saturated with step counters and calorie loggers that users abandon within two weeks. Helio Health came to us with a different thesis: the problem isn't data collection, it's that apps never translate raw numbers into actionable, personalised guidance that adapts to your actual life.

They wanted a cross-platform mobile app that could sync with any major wearable — Apple Watch, Garmin, Fitbit, Whoop — aggregate biometric data in real time, and use an on-device AI coaching layer to surface exactly the right insight at the right moment. Not a generic notification. A genuinely personal recommendation grounded in months of the user's own data.

Eighteen weeks after kick-off, Helio launched on the App Store and Google Play. Within six months it had crossed 100,000 active users with a day-30 retention rate of 68% — more than double the wellness app category average.

The Core Problem

Existing health apps collected data but didn't generate understanding. Users had step counts, sleep scores, and heart-rate graphs with no context for what any of it meant for them specifically. The result was novelty-driven engagement that collapsed after the first week, with day-30 retention across the category sitting below 30%.

Building AI that feels personal, not creepy

The technical and product challenges here were intertwined. On the technical side, wearable APIs are notoriously inconsistent — each vendor has its own data model, sync frequency, and permission model. Normalising heart-rate variability, sleep stage, and activity data across five different device ecosystems into a coherent unified schema was a significant engineering effort in itself.

The AI coaching engine presented a different kind of challenge. Health data is deeply personal, and users are sensitive to feeling surveilled. The coaching had to feel genuinely helpful rather than algorithmic. That meant running as much of the inference as possible on-device, keeping the conversational model grounded strictly in the user's own longitudinal data, and giving users full transparency and control over what data was used.

Privacy regulation added a third layer. HIPAA-adjacent data handling requirements, GDPR for European users, and App Store guidelines around health data all constrained the architecture in ways that had to be designed for upfront.

⚠️
Key Constraints
  • Five wearable ecosystems with incompatible data models and sync APIs
  • On-device ML inference required — health data cannot leave the device by default
  • HIPAA-adjacent data handling and GDPR compliance from day one
  • Sub-100ms UI response time for real-time metric updates
  • Offline-first — users must see their data without a network connection
  • iOS and Android parity with a single codebase and no compromises on feel
💡
Our Approach
  • Unified biometric schema with an adapter layer per wearable SDK
  • Core ML (iOS) and TensorFlow Lite (Android) for on-device inference
  • End-to-end encryption with user-held keys; server stores only ciphertext
  • React Native with Reanimated 3 for 60fps animations on both platforms
  • WatermelonDB for offline-first local storage with background sync
  • Longitudinal personalisation model trained per-user on-device over 14 days

A three-layer architecture: sync, store, coach

We designed Helio around three distinct layers that each had to work independently and seamlessly together. The sync layer handles all wearable communication — polling, webhook reception, conflict resolution, and normalisation into Helio's canonical biometric schema. It runs as a background task and keeps working when the app is closed.

The storage layer uses WatermelonDB, a high-performance SQLite-backed database purpose-built for React Native, giving users instant access to months of historical data with zero network dependency. Encrypted syncing to the cloud backend happens opportunistically, ensuring data is never lost even if a device is replaced.

The AI coaching layer runs a fine-tuned language model compressed to run on-device via Core ML and TensorFlow Lite. It is grounded by a retrieval system that surfaces relevant slices of the user's own historical data as context, so every response is specific to that person's patterns — not generic wellness advice.

Universal Wearable Sync
Adapter layer supports Apple Health, Google Fit, Garmin Connect, Fitbit, and Whoop. A single canonical biometric schema normalises inconsistent data models across all five platforms.
On-Device AI Coaching
A compressed language model runs entirely on the user's device, grounded by retrieval over their personal longitudinal data. Coaching is specific, private, and works offline.
Privacy by Architecture
End-to-end encryption with user-held keys means the server only ever stores ciphertext. Users can delete all data with cryptographic certainty. GDPR and HIPAA-adjacent requirements are satisfied structurally.
Offline-First Storage
WatermelonDB provides instant access to all historical data with no network required. Background sync ensures data is preserved across device replacements and never held hostage by connectivity.
Personalisation Engine
A per-user model trained on 14+ days of local data learns individual baselines for sleep, HRV, activity, and stress. Insights are always relative to your normal, not population averages.
Fluid, Native-Feel UI
Reanimated 3 and React Native Skia power 60fps gesture-driven animations on both iOS and Android. Every transition, swipe, and chart interaction was prototyped in Figma and stress-tested on real devices.
Key Decision

We chose React Native over Flutter for this project because of the depth of its native module ecosystem for wearable SDKs and HealthKit. The trade-off is slightly more bridging code, but it gave us direct access to every wearable API without waiting for community plugins. Reanimated 3 closed the animation gap entirely — users on both platforms consistently describe the app as feeling "native."

18 weeks to dual-platform launch

We ran Helio in five phases. Given the AI coaching component, we front-loaded the data and ML work so we had real on-device inference running early — not as a last-week feature bolt-on.

Weeks 1–2
Discovery & Design Sprint
Competitive analysis, user research interviews with 12 early adopters, and a full Figma prototype covering all core flows. App Store and Play Store accounts configured, wearable developer accounts provisioned.
Weeks 3–6
Data Layer & Wearable Sync
Built the canonical biometric schema, WatermelonDB storage layer, and the first two wearable adapters (Apple Health, Garmin). Established encrypted cloud sync. ML team began collecting reference biometric datasets.
Weeks 7–11
Core UI & AI Coach Integration
Implemented all primary app screens — dashboard, trends, sleep, activity, coach chat. On-device model integrated via Core ML (iOS) and TensorFlow Lite (Android). Remaining three wearable adapters completed. Internal TestFlight and Play Internal Testing launched.
Weeks 12–15
Beta, Performance & Polish
500-person closed beta. Iterated on AI coach tone and grounding based on qualitative feedback. Performance profiling reduced cold-start time from 2.1s to 0.7s. Animation pass across all screens. Accessibility audit completed.
Weeks 16–18
App Review & Launch
App Store and Play Store submissions prepared and submitted. Sailed through Apple's health data review on the first submission. Simultaneous launch on both platforms on day 126. Zero critical crashes in the first 72 hours across 8,000 installs.

Chosen for performance, privacy, and platform depth

Mobile Framework
React Native 0.74
Animations
Reanimated 3 + Skia
Local Storage
WatermelonDB
iOS ML
Core ML + Create ML
Android ML
TensorFlow Lite
LLM Backbone
Llama 3 (quantised)
Backend API
Python + FastAPI
Database
PostgreSQL 16
Auth
Auth0 + Biometrics
Cloud Sync
AWS S3 + CloudFront
Wearable SDKs
HealthKit · Health Connect · Garmin · Fitbit · Whoop
CI / CD
Expo EAS + GitHub Actions

What actually changed for users

Typical Wellness App
Generic step count goals with no personal baseline
Data siloed per wearable — no unified view
Push notifications based on time of day, not context
Population averages used as targets for everyone
AI features require constant cloud round-trips
Day-30 retention below 30% across the category
Helio
Goals set against the user's personal rolling baseline
All wearables in one unified, real-time dashboard
Context-aware nudges triggered by biometric signals
On-device personalisation model adapts to each individual
AI coaching runs fully offline, with no data sent to cloud
Day-30 retention of 68% — 2.3x the category average

Six months after launch

100k+
Active users
Crossed 100,000 monthly active users within six months of launch, driven almost entirely by organic App Store discovery and word-of-mouth.
68%
Day-30 retention
More than double the wellness app category average of 29%, attributable directly to the personalisation engine making the app feel different every week.
4.8
App Store rating
4.8 stars across 6,200+ ratings on the App Store and 4.7 on Google Play. Top reviews consistently cite the AI coach as the differentiating feature.
0.7s
Cold-start time
App loads and displays the user's full dashboard in under 700ms from a cold start, thanks to the offline-first architecture serving local data immediately.
5
Wearable platforms
Apple Health, Google Health Connect, Garmin, Fitbit, and Whoop all fully supported at launch — the broadest compatibility of any independent wellness app in its category.
0
Data breaches
Zero security incidents since launch. The E2E encryption architecture means there is no plaintext health data on servers to breach — structurally, not just procedurally.

The quality of code and architecture they delivered was beyond what we expected. We've scaled to 100k users without a single major incident. What impressed me most was how they handled the AI coaching component — they pushed back on our initial idea to send all data to the cloud and proposed the on-device approach instead. That decision is now our biggest product differentiator and our strongest privacy story. That's what a genuine engineering partner looks like.

MK
Marcus Kim
CTO, Helio Health
Next Case Study
Arcvue — Headless Commerce
View Case Study