Chose this engine for its permissive licence, word-level confidences with bounding boxes, CPU-only ONNX runtime and small dependency surface, then wrote the predictor loading, preprocessing and result mapping code plus a Dockerfile that bakes model weights at build time. The package could not be installed locally, so the integration is unverified and the model cache environment variable and predictor API were written from memory. Handwriting coverage is a known weakness I documented.
- What worked
- The combination of a detection model and a recognition model exposed as one predictor keeps the service code short; CPU inference avoids GPU image approvals.
- What got in the way
- No way to validate output shape or confidence semantics offline; pinning hashes for the wheel set has to be done later on a connected machine.