How a $30K AI build actually goes — week by week
A concrete 6–8 week ship doc, not a 28-week transformation deck. Spec call, schema, Claude integration, payments, UAT, launch. With the artifacts you'll see at the end of each week.
Prerequisites
- A clear use case (not 'we want AI somewhere')
- Someone on your side empowered to make decisions in <48h
- Access to the data the AI will read (or a path to get it)
- Budget signed off — Build $10K–$25K or Scale $25K–$50K
Week 0 — Spec call and fixed quote
15-minute intro call. NDA goes out before details. You walk us through the build: what it does, who uses it, what it integrates with, what the deadline is. We come back inside 48 hours with a fixed price and a fixed timeline.
Detailed Steps:
- NDA signed (mutual, ours or yours)
- Intro call — 15 minutes, recorded if you want
- Written one-page spec back to you within 24 hours
- Fixed quote inside 48 hours: scope, price, weekly milestones, deadline
- 50% deposit kicks off Week 1
Bring rough wireframes if you have them, even napkin sketches. Cuts spec time in half.
Expected Outcome:
Signed SOW, NDA, deposit invoice. You know the price and the date before any code is written.
Week 1 — Schema, auth, deploy
Postgres schema modelled to your domain. NextAuth or Clerk wired up. Vercel deploy live behind a password. End of week: you click a real URL and sign in.
Detailed Steps:
- Postgres schema with Drizzle ORM (Neon or Vercel Postgres)
- NextAuth (default) or Clerk (if you need MFA + org management)
- Vercel preview URL with password gate
- GitHub repo created — you have read access from day 1
- Sentry hooked up for error tracking
Push your real users into seed data this week. Test data lies. Real names, real emails, real edge cases catch real bugs.
Expected Outcome:
Live URL, working login, schema diagram in your inbox.
Weeks 2–3 — Core flows and Claude integration
The actual product. CRUD where it needs to be, AI where it earns its keep. Claude API with prompt caching enabled — input cost cut up to 90%. RAG over your data so the model answers from your records, not the open internet.
Detailed Steps:
- Core user flows built and shipped behind a feature flag
- Claude Sonnet 4.6 integration with prompt caching
- RAG pipeline: embed your data into Postgres pgvector or a vector store, retrieve on query
- Function-calling for structured output (no JSON-parsing hacks)
- Streaming UI — tokens appear as they generate, not after a 30s spinner
- Demo to you at end of week 3
If a workflow involves an irreversible action (sending money, sending email, deleting records), AI proposes — human approves. We don't ship autonomous AI on actions you can't undo.
Expected Outcome:
Working core flow with Claude integration. Working RAG. End-of-week-3 demo on your real data.
Weeks 4–5 — Auth roles, payments, admin
Multi-role access (admin, manager, user). Stripe Checkout + webhooks for billing. Internal admin panel so your team can run the platform without bothering us. Resend transactional email for receipts, password resets, alerts.
Detailed Steps:
- Role-based access control (RLS in Postgres, enforced server-side)
- Stripe Checkout, customer portal, webhooks landing in Postgres
- Internal admin panel — user impersonation, billing override, audit log viewer
- Resend transactional email with React Email templates
- Demo to your full team at end of week 5
Get your finance person on the week-5 demo. They'll catch invoice edge cases nobody else will.
Expected Outcome:
Production-shaped feature set. Your admin team can operate it without us in the loop.
Week 6 — UAT, bug fixes, launch checklist
User acceptance testing with a small group of real users. We sit on Sentry and triage every error in <24h. Pre-launch checklist: backups, monitoring, rate limits, abuse protection, status page.
Detailed Steps:
- 5–10 real users on the platform doing real work
- Sentry alerts to Slack or email
- Rate limiting on Claude/OpenAI endpoints (cost protection)
- Postgres backups configured
- Domain + DNS + SSL + email DNS records
- Launch checklist signed off
Skipping rate limits on AI endpoints is the #1 way to wake up to a $4K Anthropic bill. We default-deny.
Expected Outcome:
Zero P0 bugs in Sentry. UAT users approving. Checklist green.
Weeks 7–8 — Production, monitoring, handoff
Public launch (or internal rollout, depending on the build). Sentry alerts on errors. Daily summary email to your inbox: signups, key events, AI cost, error count. Final code handoff. 30-day post-launch fixes (60-day on Scale).
Detailed Steps:
- Production launch on your domain
- Daily summary email of platform activity
- Source code transferred (you owned it from day 1; this is the formal handoff)
- Runbook: how to deploy, how to roll back, how to read Sentry
- 30-day fix window starts (60-day on Scale tier)
- Final 25% invoice
Most clients pick up a part-time retainer at this point. Same lead, ongoing access, cheaper than re-onboarding a new dev.
Expected Outcome:
Live in production. You own everything. Daily logs in your inbox. We're on call for fixes.
Common Mistakes to Avoid
- Trying to spec everything before kickoff. The spec doc improves more from one Week 1 demo than from three weeks of meetings.
- Letting AI take irreversible actions. Always human-in-loop on send-money, send-email, delete-record paths.
- Skipping rate limits on Claude/OpenAI endpoints. One bad actor + one infinite loop = $4K bill.
- Fine-tuning before trying RAG. Prompt + retrieval gets 90% of the value at 5% of the cost.
- Not seeding with real user data in Week 1. Test data hides real bugs.
- Building an admin panel as the last step. We build it Week 4–5 so your team is operating the system before launch.