Smart Payment Solutions
Bespoke Software
Paul at Smart Payment Solutions sells POS hardware, and every product on his till interface needs a clean image. His workaround was to ask ChatGPT for ten images at a time, wait for a single collage to come back, then cut each product out by hand before loading it into his POS software. For every new product added, the whole process started again.
He wanted a private internal tool that could take a plain list of products and produce one ready-to-use image per item, correctly named and sized, with no manual cutting involved. The tool needed to stay private (it sits on a paid API key), give him control over generation cost, and let him download a whole batch in one go.
| Problem | Solution Implemented |
|---|---|
| Asking a chatbot for ten images returned one collage that had to be cut apart by hand | Built a tool that calls the image model once per product from a list, so each item arrives as its own file with nothing left to cut out |
| Product images needed to arrive at a fixed size and be named consistently before loading | Each image is produced to a fixed square size and auto-named from the product name, then packaged as a ZIP with a CSV map so a full batch loads in one pass |
| Image generation costs money and the owner needed a way to control spend | Offered draft, standard and high quality levels per run; the cheapest option costs a fraction of the highest, so day-to-day work can stay inexpensive |
| The tool holds a paid API key and should not be publicly accessible | Put the whole tool behind a passphrase gate with a durable rate limiter that survives server restarts and cannot be bypassed by rotating request headers |
| Saving several products at once could silently drop images through a race condition | Serialised the saves so parallel writes to the shared index could not overwrite each other, stopping the silent image loss that had been happening |
| A configuration rule blocked the app from reading back the generated image and still billed for it | Reworked the image handling to decode the result directly from the response, so a real generation no longer failed after the paid API call had already gone out |