The problem
An existing GPU OCR service needed a controlled front door. Separately, video annotations needed to become structured question data that people can review.
What I did
- Designed the OCR gateway and reviewed its agent-assisted implementation: authenticated requests, encrypted storage, lease-aware jobs and audit records.
- Converted video annotations into typed facts, review workbooks and multiple-choice question exports.
- Kept question generation tied to structured facts. The VQA exports were imported into the assessment platform in project 05.
Decisions
Two pipelines, no dependency
OCR and VQA share an initiative, not a runtime.
Wrap, don’t rebuild
The gateway adds access control, job handling and audit around an existing OCR service.
Questions come from facts
Generation is tied to typed facts a person can check in a workbook.
Boundaries
Service integration and data preparation around existing models; no OCR model was trained. The OCR gateway is an undeployed prototype.