Case study  /  AI e-commerce, own brand

An AI e-commerce business, built and operated end to end

AI pet portraits sell, but manual fulfillment does not scale past a hobby. I built the whole machine and I run it.

Live, revenue-generating
  • Serverless AWS pipeline
  • Human-in-the-loop review
  • 584-image eval corpus
Petropolitan case study card

The engagement

Challenge

The AI is the easy part; running a business on it is not

Turning a customer photo into a printed canvas means analyze, generate, review, and fulfill, over and over. Doing it by hand does not scale, and letting AI ship unreviewed portraits to paying customers is not an option. It needs a pipeline that keeps a human in the quality loop.

Action

The whole machine, one operator

A serverless AWS order pipeline (Lambda, Step Functions, DynamoDB, S3, SES, API Gateway) automating Etsy intake through Printful fulfillment around a human review stage; a studio orchestrating Claude for photo analysis and prompt generation with OpenAI and Gemini image models for output; an iOS app on the same backend; and a labeled eval corpus feeding a quality gate.

Result

A live business, with AI quality measured not assumed

A revenue-generating business run by one person. The quality is not vibes: 584 scored portraits and a shadow-mode quality gate that measures whether the AI output is actually good enough before it reaches a customer.

How it is built

One photo to a finished product

Etsy order
intake and customer photos
serverless pipeline→
AI studio
analyze, generate, human review
on approval→
Printful
printed and shipped

A human review stage sits between generation and fulfillment. No portrait ships unseen.

  • 01Order pipeline. A Step Functions state machine drives intake, photo collection, generation, review, and fulfillment.
  • 02AI studio. Claude analyzes the pet photos and hardens a style-specific prompt; OpenAI and Gemini image models render the portrait.
  • 03Quality loop. A 584-image labeled corpus and a shadow-mode quality gate score likeness and anatomy, the input that moves the gate from advisory to enforcing.
  • 04Native client. A SwiftUI iOS app consumes the same backend, with a refine-and-version loop.
  • 05Human in the loop. An operator approves a favorite from a small gallery before anything is pushed to print.

What it proves

I do not just build AI features. I run a business on them, and I measure whether the AI is actually good enough.

See the rest of the work →

Petropolitan is my own brand. No customer data is shown; sample imagery uses my own pets.