A Relational Household CRM, Built in Public
A relational household CRM for families. Inbox-driven capture, human-curated.
Built with
What
A relational household CRM — family, properties, assets, kids, contacts, documents, and history, structured the way a business runs its book of accounts. AI handles capture from the inbox; humans curate.
Why I’m building it
After a decade implementing CRM systems for businesses, the gap on the household side is hard to miss. Families run on the same primitives a business does — people, properties, assets, documents, history — but the tooling assumes a company on the other end. The bet is that any household with the inclination to keep a single source of truth eventually wants one, and that the right shape for it is relational, not a folder tree. I’m building it in the open because the harder problem isn’t the product — it’s pushing on AI-augmented product development as a methodology that actually ships.
How it’s built
A six-agent governance harness — PM, design, frontend, backend, extraction, and trust-ops — with hard rules enforced in CI. The defining one: every locked extractor version is eval-gated, with per-field regression against a stored baseline blocking the commit. The interesting work isn’t the extraction — it’s earning trust at the trust boundary, where the raw email body is never persisted, medical mail is metadata-only, and the OAuth scope stays read-only forever.
Stack at a glance
- Next.js 15 with React 19 — App Router, Server Actions
- Tailwind v4 + shadcn/ui — design tokens via
@theme - Supabase + Drizzle — Postgres + Auth + Storage with RLS, typed schema and migrations
- Inngest — durable functions for the inbox-ingest pipeline
- Anthropic Claude via the Vercel AI SDK — extractors at the trust boundary
- Resend + Sentry — transactional mail and observability
- Vitest with pglite + Playwright — RLS-aware unit specs and end-to-end smoke
- Hosted on Vercel
What this build is teaching me
Solo product building with an AI-native stack is a different shape than it was five years ago — eval gates do for prompts what type systems do for code, and the unit economics of a one-operator team shift accordingly. The discipline this project demands is mostly about deciding what the trust boundary is, then writing the regression tests for crossing it.
