Naru AI

Naru AI was an iOS family health journal: symptom logging with severity, mood and photos, a medication tracker with reminders, an AI triage chat, and shareable reports — scoped to multiple family profiles. It went from first Swift commit through App Store review to a paid auto-renewing subscription, metered at roughly two million AI tokens per user per month. Pricing that tier forced me to treat per-user AI cost as unit economics, not an API bill.

The architecture was local-first, not a thin API client. SwiftData is the source of truth on the device, with a background sync manager doing optimistic updates; an edge function ingests each entry, runs LLM extraction to turn free text into structured health events, and writes embeddings to Postgres with pgvector so the app can answer questions across months of entries.

The model never got raw database access. Health data was exposed through an MCP server over stdio with narrow, purpose-built tools — a retrieval boundary that keeps the model’s reach explicit and auditable, running in a shipped product rather than a demo.

Then the measurement came in, and usage did not justify running the backend. The backend was shut down and the records kept — a call made on evidence, which is exactly what market testing is for. What the product taught carried forward: its retrieval boundary became StemBuddy’s review surface, and its cost discipline became SEAM’s deploy gate.