How I'd build Architect AI
I'd reach for Next.js and FastAPI: Next.js handles the document editor and templates, FastAPI runs the Claude integration and document generation in the background. Postgres holds the tenant data and document history. I'd estimate 160-180 hours for an MVP that handles specifications, proposals, and change orders, with compliance notes shipped behind a professional-liability banner that transfers legal responsibility to the firm.
Day-by-day plan
- Day 1-2: Provision Postgres schema (tenants, users, documents, templates), Clerk auth, multi-tenant query guards. FastAPI scaffolding with request logging.
- Day 2-3: Wire Stripe billing for three tiers (Pro, Business, Enterprise). Webhook handlers for subscription state changes. Grace-period handling.
- Day 3-4: Claude API integration layer. Build template system: specs, proposals, change orders, RFI drafts. Generate PDFs with Weasyprint. Async job queue with Celery and Redis.
- Day 4-5: Next.js document editor with Tiptap. Onboarding flow: firm setup, template selection, first document generation, preview and download.
- Day 5-6: Document version control and audit log. Users can see all changes and who made them. Restore to any prior version.
- Day 6-7: Compliance disclaimers for code compliance notes. Terms and professional-liability language. Require explicit acknowledgment before generation.
- Day 7-8: Error recovery, rate-limiting on Claude API, retry logic. E2E tests for document generation. Deploy to Vercel and Railway.
What's hard about this build
The legal minefield is real. Architecture firms carry professional liability insurance, and compliance notes can open them to lawsuits if they're wrong. I can't ship a compliance-notes feature without legal review, which means this isn't a launch feature. That shrinks the value prop significantly. Second: AIA standard forms (B101, A101, G701) are copyrighted and closely guarded. I can't bake them into the product - lawyers will kill the deal. Instead, I let users upload and customize their own templates, which is slower to onboard but defensible. Third: Claude hallucinates. If a proposal generator confidently makes up a material requirement or exclusion, that's a real problem. I'd need human review workflows built in, which adds friction.
What's fast because of AI
Claude compresses weeks of scaffolding into hours. I prompt for the Postgres schema, multi-tenant guards, and Stripe webhooks. Claude Code generates 80% working code that I refine. Test writing: I describe the document-generation flow, Claude enumerates edge cases and writes Jest and pytest specs. Copywriting: onboarding screens, email templates, compliance disclaimers - all drafted by Claude with minimal revision. Debugging: hallucinations in generated documents are caught by asking Claude to enumerate what could go wrong before generation, which narrows the failure modes. API design: I iterate on the document-generation endpoint by asking Claude for design critiques, which saves a round-trip of thinking.
How I'd hand it off
I'd record a 20-minute Loom walkthrough of each feature: onboarding, template selection, document generation, payment setup. Leave a runbook in Linear covering deployment, Stripe/Clerk/Anthropic credential rotation, scaling the Celery queue, and monitoring Sentry alerts. I'd do a 30-day on-call rotation where I'm the pager contact for production issues. Transfer all service credentials to your account. Leave the Linear board with known issues, the roadmap (compliance notes pending legal, better AIA template support, API for integrations), and a note on what I'd build next if the product survives customer traction.