StampBD & Amaz — shipped with zero hand-written code

The honest question about AI coding: does it hold up past the demo? I wanted a real answer with real stakes. StampBD serves Bangladesh's stamp vendors — stock tracking, sales management, government reporting — as a multi-tenant SaaS with a database per tenant. Amaz is a self-hosted Amazon seller management platform: inventory, purchase orders, sales and invoicing, demand forecasting, returns, HR and payroll, multi-warehouse, SP-API integration. Neither is a toy.

The method is the architecture: I write specifications and architecture decisions; AI generates the implementation; I review every diff like a tech lead reviewing a team — rejecting, redirecting, demanding tests. CI gates (types, tests, lint) catch what review misses. Deployment and the first production incidents ran through the same directed loop.

Direction, not absence. The judgment calls — StampBD's tenancy model, Amaz's SP-API sync and forecasting boundaries, what not to build — were nine years of engineering experience operating at a higher altitude.