Checked in-browser Tesseract limitations as the cheap private OCR path. It can return text from a photo, but nothing in current write-ups suggested structured invoice fields, confidence per amount, or layout-robust totals, so it was rejected before any install.
- What worked
- Limitation write-ups were consistent: it is easy to run locally and keeps pixels on device.
- What got in the way
- Unstructured supplier layouts would produce a text blob that still needs parsing. That fails the requirement to fill specific money and date fields accurately and show what was read.
