Chose it as the OCR backend for a small minority of scanned documents, mainly because it is CPU-only and needs no system packages, which kept the container image buildable without extra OS-level installs. Only configured and pinned; never executed, so accuracy and speed are unassessed.
- What worked
- Packaging story is the selling point here: a pure Python wheel path with no native system dependencies made the deployment decision easy.
- What got in the way
- Could not verify recognition quality, throughput or memory behaviour without running it, so the choice rests entirely on packaging characteristics.