Hypothesis

Problem

Wellness apps collect plenty of data but never answer the question users actually care about — why do I feel this way? People end up with dashboards full of metrics and no idea which of their habits is responsible.

Approach

1. 01Designed a four-step flow — define an outcome, list candidate activities, generate a hypothesis, track it over time — so the app teaches the scientific method by structure rather than instruction.
2. 02Used GPT-4o-mini with Gemini as a fallback to turn vague goals into neutrally worded, testable predictions, deliberately avoiding language that nudges the user toward a result.
3. 03Integrated the WHOOP API over OAuth to pull 180 days of recovery, strain, HRV, resting heart rate, and sleep performance, persisting tokens so the connection survives app restarts.
4. 04Wrote an enrichment step that converts WHOOP metrics into binary features (high recovery, good sleep, high strain) and feeds them into a regression alongside self-reported activities.
5. 05Kept every API key server-side behind an Express proxy, then ported that proxy to Cloudflare Pages Functions — including an edge-native rate limiter, since an in-memory one cannot survive Workers isolates.

Outcome

A working iOS, Android, and web build that reports coefficients like “Good Sleep (>85%): +1.5 (strong positive)” next to a user’s own habits, plus a generated one-week plan and a predicted-vs-actual scatter plot.

• Local-first storage — no user wellness data leaves the device.
• Clean git history with no exposed secrets; all credentials via gitignored env files.