How I'd build Nurture AI
I'd reach for Next.js on the frontend with FastAPI on the backend, Postgres for the database, Celery + Redis for the job queue, and Stripe for billing. The email sending goes through Resend, and I'd wire the Gmail and Microsoft Graph APIs for monitoring opens and routing replies. The agent itself runs on Claude API via the Python SDK. Rough estimate: 180-220 hours to production, which puts us at 3-4 weeks at full throttle. That's with tests, auth, and a working onboarding flow.
Day-by-day plan
- Days 1-2: Auth schema (NextAuth.js), multi-tenant database model, and Postgres migrations. Stripe customer table linked to tenants.
- Day 3: Stripe integration: product setup, payment webhook handling, subscription state logic across the three tiers.
- Days 4-5: Resend email sending wrapper and basic send tracking. Build the mailbox connection flow (user authenticates Gmail/Outlook via OAuth).
- Days 6-7: Gmail API integration for pull-based monitoring: open tracking via pixel logic, reply ingestion into Postgres, unsubscribe handling.
- Days 8-10: Agent scaffolding in FastAPI. Build the Claude API bridge, prompt templates for sequence draft and reply classification. Test end-to-end with synthetic data.
- Days 11-14: Onboarding UI. List upload (CSV parser), offer description intake, template selection, sequence review before first send. Wire to the agent.
- Days 15-16: Monitoring dashboard. Reply inbox, open/reply rates, sequence status view. Basic Celery job status tracking.
- Day 17: Runbook, error handling, rate-limit logic for Gmail/Outlook APIs. Docker setup and Railway deployment.
- Days 18-19: Testing, edge cases, and documentation. Permission model for team members sharing a workspace.
What's hard about this build
The email deliverability piece is the silent killer. Gmail's authentication requirements (DMARC, DKIM, SPF alignment on customer domains) and Outlook's sender reputation scoring mean that agent-generated copy still lands in spam if the infrastructure isn't right. That's a legal and technical surface I'd need to audit early: we're responsible if sequences get flagged, and we need clear documentation on what customers must set up on their own domain. The Gmail and Outlook API integrations are also fragile around permission scoping and token refresh - I'd build a robust retry and monitoring layer there. Reply classification via Claude API also has a latency tail; if a customer gets 500 replies in 4 hours, we need Celery to batch process without blocking the UI. The other risk is offer-to-list fit: if the agent writes solid copy but the customer's list is stale or misaligned to their offer, sequences will underperform and they'll blame the tool. That's a support and churn problem, not a build problem, but it shapes how I instrument observability.
What's fast because of AI
Claude compresses the agent backbone from a week to 3 days. I'm using Claude for the sequence drafting, the reply classification, and the offer-to-list diagnostics. The standard flow would be: define state machines for each step, handle 15 edge cases manually, debug permutations in production. Instead, I write the core prompt, Claude handles the nuance, and I test against known failure modes. Edge-case enumeration - what happens when a reply is sarcastic, or a name is ambiguous, or the unsubscribe link is in the footer - Claude covers 80% in the first pass. Scaffolding the FastAPI + Next.js boilerplate also compresses by 2-3 days because I'm not hand-writing auth flows and CRUD patterns. I'm using Claude for test generation too: once I define the happy path, I ask for tests covering schema validation, permission checks, and rate limits, and I refine from there rather than write from scratch.
How I'd hand it off
I'd record a Loom walkthrough of the build: here's where the agent logic lives, here's how to monitor Celery jobs, here's the Stripe webhook logs. The runbook covers OAuth setup (you need Gmail and Outlook service accounts), Rails migrations, environment variables for API keys, and the deployment pipeline. I'd set up a 30-day pager rotation where I'm on-call for critical issues (sequences stuck in Celery, email delivery failures, Stripe auth edge cases). All credentials go into a shared LastPass vault or 1Password. The Linear board stays open so you can see what's queued. I'd also leave a prompt template library and a test suite so the next engineer can iterate without guessing.