Running the vision models without a GPU and without training data seems impossible - until you run them as Lambda containers.
I took two vision models, put them into Lambda containers and made them process images automatically. The images go into an online community catalog of collectible AWS Builder Cards - buildercards.dev.

This is how it works:
- Anyone takes a photo of a card and uploads it to the web
- A vision model separates AWS Builder Cards from everything else
- A second vision model removes the background, and math tools inside the same Lambda straighten and crop it
- A multimodal model from Amazon Bedrock extracts the text and stores it
- The data is committed to a GitHub repo, from where Amplify pulls it and generates and deploys the webpage

In the whole process there is just a single human approval. Both vision models run in Lambda containers without GPUs, and without me training or fine-tuning anything, which makes the inference extemly cheap.
This is a demonstration that not all jobs require a "PhD model" or extermly powerful hardware to do the job.