Nabd (نبض)

The problem. A WHOOP band produces months of recovery, sleep and strain numbers, and reading them well takes context. I wanted a private dashboard, in Arabic and English, that explains the data in plain language, finds the wearer's own normal ranges and correlations, and turns them into a two-week plan, without raw health rows ever leaving the owner's machine.

The hard part. Plain-language explanations usually mean sending raw data to a model. Nabd only ever sends computed summaries, never rows: a privacy guard is the last check before any request leaves the machine, a "Show what is sent" control displays the exact payload, and a monthly token budget caps spend. The analytics also refuse to present a correlation as a cause, which shaped both the maths and the copy.

The approach. Four one-way layers run from ingestion to analytics to API to app. The analytics package is pure TypeScript with no I/O and is the only place a number is computed. Two ingestion paths, the WHOOP CSV export and the WHOOP API on a scheduled sync, write through one upsert. The app is Nuxt 4 on SQLite via libSQL and Drizzle, styled with Tailwind logical properties and self-hosted IBM Plex Sans Arabic.

The outcome. Nabd was built in phases, each closed by a written gate, with a decision log of more than 360 entries. Contrast is enforced by a test, and every screen is captured in Arabic and English, light and dark, at 390 and 1280 pixels. The Phase 14 gate report recorded 3,795 unit tests and 1,120 end-to-end tests passing, and the Phase 13 gate recorded Lighthouse accessibility 100 on the four main screens, with performance between 87 and 94. The screenshots here use synthetic data.