Wishdeal Factory · Storefront
Operator interview · $75/hr · Roll Digital's seat
← Back to ManufactureAI

How Caleb would build ManufactureAI.

First-person from one of our chief operators. What he'd ship and how, AI-amplified. Stack, hour estimate, day-by-day plan, the parts that are hard, and the handoff. Synthesized from the agent spec.

How I'd build ManufactureAI

I'd reach for Next.js on the frontend, FastAPI for the backend, Postgres for multi-tenant data, Stripe for billing, and Resend for transactional email. The quoting engine itself - the core of this product - would be a thin Python service that ingests job histories, learns shop-specific pricing patterns, and outputs bid recommendations via the API. Rough estimate: 400-420 hours, which puts us comfortably inside the $32k budget with room for deployment and documentation.

Day-by-day plan

Day 1: Provision auth schema with Clerk or Auth0, set up multi-tenant isolation at the database layer, and scaffold the main Next.js app structure. Day 2: Wire Stripe billing across three pricing tiers, set up webhook handlers for subscription lifecycle events. Day 3-4: Build the customer onboarding flow - email verification, shop details form, initial quoting template import. Day 5-6: Build the core quoting interface and the FastAPI backend that takes job history as CSV input and surfaces pricing recommendations. Day 7: Integrate Stripe's white-label quoting setup - allow customers to invite their team members and set permissions. Week 2: Build the analytics dashboard showing save rate on quoted jobs, conversion tracking back to actual bids. Week 3: Spin up the per-shop training data ingestion - let shops upload their historical job data, label it, and retrain the quoting model. Week 4: Testing, edge cases, database migration scripts. Week 5: Deploy to staging, run through the GTM flow as a customer would see it, iterate. Week 6: Production deployment, monitoring setup, Datadog alerts on quoting latency.

What's hard about this build

The real risk isn't the tech - it's the data problem. Quoting is tribal knowledge. A shop's pricing reflects decades of material supplier relationships, machine utilization patterns, labor cost structure, and how the owner thinks about margin. An AI that doesn't learn this quickly becomes a liability. We need the onboarding to extract enough historical job data in the first week to train a model that's accurate within the shop's pricing variance. If we miss that window, churn is high. The second risk is incumbent lock-in. Most shops run JobBoss or Shoptech or E2 Shop, and their entire quoting workflow lives there. We're not replacing the ERP - we're sitting alongside it as a "second brain" for better pricing. But switching cost is real labor, not just money. The third risk is the trust barrier. AI isn't a selling point to a skeptical shop owner; a saved lost bid is. Demos have to show a real saved job, not a list of features.

What's fast because of AI

Claude compresses weeks into days on scaffolding - the multi-tenant schema, API routes, middleware patterns. I prompt for the entire database migration strategy and get a working schema in an hour instead of building it from scratch. Test generation is huge: I write one test case for quoting logic and Claude enumerates edge cases - material surcharges, rush fees, setup costs, minimum orders - then generates parametrized tests. Copywriting: UI microcopy, email sequences for the GTM, onboarding docs. Instead of writing ten emails myself, I describe the customer journey and Claude generates options I shape. Debugging is faster too. When a quoting calculation is off by 3 percent on a specific material type, Claude helps me trace through the logic, spot the rounding error, and write the fix.

How I'd hand it off

I'd record a Loom walkthrough covering the admin panel, customer onboarding workflow, and how to run a quoting job end-to-end. Runbook goes in a private wiki: how to backup the Postgres database, how the Stripe webhook handler works, how to scale the quoting service if throughput spikes. I'd do 30 days of pager rotation - any tier-2 issues come to me first, then I hand them over with context. Tools and credentials: GitHub repo access, Linear for bug tracking, Stripe and Postgres credentials stored in 1Password with instructions for rotation.

Hire Caleb to build this for you.

ManufactureAI is available to own for $200 flat. Or pay $75/hr for a Roll Digital chief operator to build it for you, AI-amplified.

See pricing →