Reading UPD invoices from phone photos: the real bottleneck is 44 GB of RAM

On October 3, 2026, a developer published a post on DEV Community about a tool he built with a coding agent for his wife, an accountant.

The task was deliberately narrow. Take a phone photo of an UPD, the standard Russian invoice form, read the VAT rate from column 7, group the rows by that rate, sum columns 8 and 9, and check those sums against the document total. The pipeline runs photo → locate the three columns → read the numbers → group by VAT rate → verify the totals. His wife had been doing this by hand, which is what prompted him to start building the tool.

What he expected to be the hard part did not turn out that way. OCR, he writes, was the easiest stage of the whole exercise. The real obstacle was memory: the pipeline needed 44 GB to run. That is the claim the post leads with — reading three columns of numbers off an invoice takes 44 GB, and the OCR model is not the problem.

The account stops there. The post does not say which stage of the pipeline accounts for the 44 GB, nor does it describe any attempt to bring the memory footprint down or to replace the approach with something lighter. For now the write-up ends at the diagnosis rather than at a fix.