Order Pipeline Toolkit

Why I Built This

Fulfillment was slow in the gaps between steps

A personalized order passes through the same steps: buy a shipping label, print what goes in the box, find the customer's photo, turn that photo into something a laser can engrave. None of those tasks is hard. What made the day slow was that they lived in different places and had to be done in order, by someone who remembered the order. After printing, someone had to open every photo order and download the customer's photo by hand, then edit those photos one at a time. People waited on downloads, and checked in to ask whether a batch was still running. Every new hire had to learn the whole sequence before they could fill a single order.

When I opened the existing printing program, I found two problems nobody could see from the screen. It never checked whether an order already had a label, and it did not compare shipping options against the delivery speed the customer paid for. I built this system so the sequence lives in the software instead of in someone's head: press one button, and the right label and contents sheet print while the customer's photos download into today's folder, ready for the photo generator. A new hire can now run a batch without learning the old sequence first.

At a glance

What changed for the people using it

The old printing program next to my rebuild, one row per task. The sections below show each change on screen.

How It Works

The structure of the system

Two programs, one pipeline. The Print Assistant pulls orders from the shops, buys or reuses the right label and prints. When an order needs the customer's pet photo, the program downloads that photo into a dated folder, and the AP Photo Generator turns it into engraving-ready line art.

Program 01 · Print Assistant

Before and after: what I improved

Same program, same services, two versions. Before is v1.4, the version the team used before my upgrade. After is v2.10, my rebuild. I kept the parts that worked, the ShipStation and Veeqo requests and the order they run in, and rebuilt what people touch and what could cost money.

1 · The controls: from stacked forms to three numbered steps

Before · v1.4

After · v2.10

2 · Reprinting: from the shipping system to three buttons

Sheets jam, labels smudge, an order changes. Before, fixing any of that meant leaving the program: log in to the shipping system, find the order, download the label, and print it separately. v1.4 could only reprint the contents sheet. In v2.10 one order field and three buttons cover every case, and reprinting a label never buys a second one.

Before · v1.4

After · v2.10

3 · While it runs: from a scrolling log to a finish time

Before · v1.4

After · v2.10

Program 01 · Error messages

4 · When something goes wrong, in the program's own words

⚠️ ShipStation: Order Not Found: [order #]

⚠️  Order [order #] was NOT found under Amazon.
    • Is it an Amazon order? (Etsy/Shopify orders won't be here, pick that store instead.)
    • Double-check the order number.
    • Confirm your API keys are set (Settings ▸ API Keys).

💳 Shipping Label Requested (Order: [order #])
✅ Veeqo Label Purchase Succeeded; label price: ...

   🚚 Label bought: USPS Priority Mail  ← requested carrier + service
♻️  Label already bought for [order #], reprinting the SAME label (no new charge).

Program 01 · How I built it

How each Print Assistant function works

v2.10 is a Python desktop program for Windows, built with Claude Code and ChatGPT. It is packaged into a single app, so a new computer runs it without installing anything. Open the list below for what each part does and how I built it.

Program 02 · AP Photo Generator

Step by step: from customer photos to engraving files

Personalized pet pieces are engraved from a line drawing of the customer's dog. Those drawings used to be edited by hand, one order at a time. The generator is a small web app that runs on the work computer and sends photos to the OpenAI Image API in parallel. Here is how a batch works, screen by screen. These are the real screens: customer photos are blurred, and order numbers and the Drive folder ID are masked.

1. 01

   Add the photos

   Drag in the folder the Print Assistant just made, like 09.16.2026 Amazon, or click to choose files.

   • JPG, PNG or WebP, up to 20 MB each
   • Up to 100 photos per batch, nested folders included
   • Each photo gets a card right away, and any card can be removed with ×

   How I built it: Add the photos

   The generator is a small Flask web app that runs on the work computer and opens itself in the browser. Dropped folders are read recursively in the browser, which is why the Print Assistant's 09.16.2026 Amazon folder can go straight in. The server checks every file for type, size and a readable image before anything is sent out, and converts unusual formats for the model without touching the original.

   

2. 02

   Choose the output

   Settings are chosen once for the whole batch.

   • Quality: Draft, Balanced or Final
   • JPG size: from 300 px squares up to 2048 px
   • Additional instructions, optional, for a specific detail like an ear fold
   • Google Drive: pick a folder and save into a dated subfolder
   • After generation: upload to Drive and/or download everything as a ZIP automatically

   How I built it: Choose the output

   The server only accepts the models, qualities and sizes on the list. The prompt is built from a locked template, and the optional note is added after the rules, capped at 2,000 characters, so it can refine a detail but never override the locked rules.

   

3. 03

   Generate, all at once

   One click starts every photo together as its own request, instead of waiting for them one by one. This parallel run is where the 95% cut in editing time comes from.

   • A progress bar counts finished and in-progress portraits
   • Finished cards unlock Download JPG while the rest keep drawing
   • Stop all cancels running jobs and keeps what is already done

   How I built it: Generate, all at once

   Each photo is its own OpenAI Image API request run in background mode, so the server hands back a job ID right away instead of holding the connection open. The page checks every job on a schedule spaced to stay within a request budget, and retries dropped connections with backoff. A retry only re-reads a job and never pays for a new one. Each result is checked as a real JPG and resized to the chosen size.

   

4. 04

   Compare side by side

   Click any finished portrait to open the compare view: the original photo on the left, the generated JPG on the right.

   • Zoom both images together to check eyes, ears and nose at the same spot
   • Previous / Next steps through the whole batch without closing

   How I built it: Compare side by side

   A single zoom value drives both panes, so the two images always stay at the same scale. The arrow keys also move between photos.

   

5. 05

   Redo just the one that is off

   If one drawing misses, only that photo is sent again, with its own settings and its own note.

   • Change quality or add a redo prompt for that photo only
   • The previous JPG stays until the new one succeeds
   • Other photos can be redone while this one runs

   How I built it: Redo just the one that is off

   A redo is a separate request with its own settings. The new JPG replaces the old one only if it succeeds, and the dialog states that it is a paid request before anything is sent.

   

6. 06

   Save and hand off

   When the batch is done, the files go where the engraving work picks them up.

   • Download all as one ZIP
   • Upload to Google Drive: each original and its generated JPG, as a pair
   • Restart / new batch clears the page for the next set

   How I built it: Save and hand off

   The ZIP is assembled in the browser by a small ZIP writer I wrote, so nothing is uploaded to make it. Drive uploads go through the Google Drive API with a Google sign-in: the app creates today's dated folder inside the chosen folder and uploads each original with its JPG as a pair.