The last 30% of an AI-built SaaS.

The last 30%

What separates a demo from a product

1. 01Billing that cannot leak money
2. 02Tenant isolation that cannot leak data
3. 03A deployment that can be rolled back
4. 04A privacy posture a European buyer's procurement team will accept
5. 05A release discipline that lets you ship on Friday without spending Saturday on fixes

None of it shows up in a demo. All of it shows up in a procurement questionnaire, a failed renewal or a support call.

Curriculum

Eight weeks, one product: yours

1. 01The Conductor loop

   Specify, plan, build, review, record. A PRD an agent can execute, standing orders in CLAUDE.md, and review without reading every line.

2. 02Stack and bootstrap

   Next.js, Supabase, Stripe and the Claude API, chosen once. Three environments and secrets handled properly before the first feature.

3. 03GitHub, CI and tests

   PRD to epics to issues. The pull request as the unit of work, branch protection, and a test pyramid the agent writes and you enforce.

4. 04Tenancy and row-level security

   Organisations as tenants, RLS on every table, and the two-tenant isolation test that runs on every push.

5. 05The feature build loop

   Vertical slices, a fixed build prompt, and the predictable ways an agent goes wrong.

6. 06Billing

   Verified, idempotent Stripe webhooks, entitlements derived from subscription state, and every renewal and failure state tested.

7. 07AI in the product

   Metering, budgets and ceilings, prompts as versioned code, and an evaluation set that catches quality drift.

8. 08EU-ready and released

   GDPR without a lawyer for v1, a security baseline, observability, release reviews and smoke tests, and the first customer.

Who it is for

Built for people who specify

• Founders and consultants with a validated idea, no engineer, and a budget measured in evenings rather than salaries.
• Product, operations and finance people who can specify precisely and want to ship without a dev team.
• Developers who are new to running agents and keep ending up with sprawling, untested codebases.

It is not for people who want to learn to program, and it is not a Claude Code tutorial. It assumes you can already get an agent to build a feature.

Taught by Philip Rutovitz, who spent twenty years running regulated operations in trust and SPV administration before building four software products by directing AI coding agents. The method comes from those builds: the MVP took a week, and the last 30% took three months.