EZ Health

Apple Health analytics with an AI training coach. Import your iPhone's full Health export (or stream it daily from an iOS automation) and get iOS-style dashboards for heart rate, HRV, sleep stages, activity patterns and workouts, plus a coach that plans running and strength sessions from your actual recovery data.

Live demo: ez-health-six.vercel.app (generated demo data; the AI chat is switched off on the public demo)

Screenshots use the built-in generated demo data (npm run demo:import), not real health records.

Highlights

• Imports of any size, even on serverless. export.zip is unzipped and parsed as a stream (SAX) in a Web Worker, and only the compact parsed records are uploaded, in batches of 3 MB at most, under Vercel's 4.5 MB request limit. The same parser runs the CLI import, where a 1.38M-record export takes about 58 s while memory stays around 32 MB (it was 678 MB before source strings were interned).
• Idempotent bulk writes. Upserts use Postgres unnest() + ON CONFLICT, thousands of rows per statement, so re-importing the same file or re-sent webhook windows update data instead of duplicating it.
• Timezone-correct analytics. Every sample keeps its original UTC offset, plus precomputed local date, hour and weekday. "Most active hour" and "bedtime" mean the time on your clock, even across travel.
• Multi-device de-duplication. iPhone and Apple Watch both log the same activity. Data is stored per source and merged at query time, for example the best-measured night for sleep and the per-hour maximum for active energy.
• Explainable scores. Readiness, sleep score, consistency and the 7/30/90-day trend verdicts are deterministic functions with documented thresholds. A change only counts as a trend when it exceeds normal day-to-day variation.
• Grounded, private AI coach. The model receives a compact text summary computed by the same code as the charts, so the advice and the dashboards never disagree, and raw samples never leave the server. Provider-agnostic: OpenAI or Gemini.
• Accessible charts. Palettes are checked for color-vision deficiency and contrast in both themes, every chart has a "Show as table" view, and dark mode has its own validated colors rather than inverted ones.
• English and Hebrew, with a full right-to-left layout. A language switch, or the browser's language, sets the copy, the 24-hour clock, dates and units, and the AI coach answers in that language. The layout mirrors through logical CSS properties, while charts keep time running left to right.

Screenshots

Summary	Summary (dark)
Heart & Breathing	AI Coach
Workouts & Trends	Import

Hebrew (right to left)

Sleep & Recovery in Hebrew	Summary in Hebrew, iPhone

Dashboards

Page	What it shows
Summary	Readiness, sleep score and the latest reading of each metric
Heart & Breathing	Resting vs workout heart rate, HRV, respiratory rate; 7/30/90-day trend verdicts
Sleep & Recovery	Last night's stage timeline, bedtime/wake schedule, stages per night, sleep score, consistency
Workouts & Trends	Fitness & recovery trend evaluator (7/30/90 days), weekday × hour activity heatmap, side-by-side workout comparison
AI Coach	Rule-based readiness + a chat coach (OpenAI or Gemini via the Vercel AI SDK)

All scores are documented where they're computed: src/lib/analytics/ (trend.ts, sleep.ts, readiness.ts) and src/lib/queries/evaluation.ts. The coach's summary is built in src/lib/coach/snapshot.ts.

Architecture

flowchart LR
  A["iPhone Health export<br/>export.zip / export.xml"] -->|"/upload"| P["In-browser Web Worker:<br/>streaming unzip + SAX parser"]
  B["Health Auto Export app<br/>(daily automation)"] -->|"POST /api/health-ingest<br/>x-api-key"| Z["Zod validation<br/>+ normalisation"]
  P -->|"POST /api/import<br/>batches ≤ 3 MB, zod-validated"| S["Idempotent bulk upserts<br/>(unnest + ON CONFLICT)"]
  Z --> S
  S --> DB[("PostgreSQL<br/>Prisma 7")]
  DB --> Q["Query + analytics layer<br/>trends, scores, de-dup"]
  Q --> UI["Server-rendered dashboards<br/>Recharts"]
  Q --> C["Coach snapshot"] --> LLM["OpenAI / Gemini<br/>Vercel AI SDK"]

Stack: Next.js 15 (App Router, Server Components, route handlers), React 19, TypeScript, Tailwind CSS v4, Motion, Recharts, Lucide, Prisma 7 with the pg driver adapter, PostgreSQL (local Docker, Supabase or Prisma Postgres), zod, saxes, jszip, Vercel AI SDK.

This is the beginning of the README. Read the rest at the original. Read the full README ↗