Area C—Data Foundation
RedBrick Annotation ToolKit
Annotation operations as code
One config and one command per annotation project — every operation audit-logged for regulators.
- versioned project configs
- 157
- versioned project configs
- operations audit-logged
- 100%
- operations audit-logged
01/Problem
Regulated ground truth means many annotation projects, each with its own taxonomy, multi-reader protocol and export. Doing it by hand in the vendor UI was slow, error-prone and — critically for an FDA-bound pipeline — unauditable.
02/Approach
A declarative configuration plus a single command drives project creation, upload, label import, reviewer assignment and export across X-ray, CT and MRI, for classification, polygon, landmark and voxel segmentation. It automates multi-reader protocols directly — including “every annotator labels every case” for inter-reader agreement — and imports model predictions so annotators correct rather than start from scratch.
03/Outcome
- 10 product lines, 157 versioned project configurations, 31 workflow notebooks.
- 100% of operations logged with operator, timestamp and resolved configuration.
- Pre-label import a major contributor to the 50% annotation-effort reduction.
04/My role
- Evaluated and onboarded the annotation platform as part of a vendor consolidation that cut tooling costs 20%.
- Required protocols be version-controlled configuration so consensus setups behind regulated models are reproducible under submission scrutiny.