Challenge
A national security AI program needed large volumes of accurately labeled imagery and video to train and refine computer vision models. Algorithm vendors worked in sprints of 90 to 120 days, so datasets had to arrive on time and at a consistent quality. Labeling errors compound through an AI pipeline, degrading model performance and forcing costly rework, so accuracy mattered as much as speed. The work also called for cleared personnel and strict data handling, and the staffing arrangement changed partway through.
Our Approach
CEdge, through the Simpack-Edge joint venture, supplied the labeling workforce and managed delivery as a subcontractor on a larger program.
Labeling to the customer’s standards
Analysts drew bounding boxes and attribute labels around objects of interest such as vehicles, structures, and equipment, following the customer’s ontologies and evolving labeling guidance. The work spanned several sensor types: electro-optical, synthetic aperture radar, full motion video, and horizontal motion imagery.
Quality built into the workflow
Every dataset passed through multi-level review before submission, combining human-in-the-loop checks with automated validation. Structured feedback between annotators and quality reviewers kept accuracy high, and personnel on improvement plans were coached above the quality threshold.
A workforce transition with no gap
When the staffing structure changed, the joint venture moved the existing workforce into its own organization without interrupting operations or billing. We mirrored CEdge’s HR, payroll, and compliance policies, completed offers, benefits, security training, and timesheet setup without a break, and secured favorable HSPD-12 adjudications for all personnel within 60 days. A dedicated management team stayed in place throughout.
Going beyond the agreement
We also built an integration that let the prime pull labor and timekeeping data directly from our systems, improving its visibility and simplifying reporting, even though the agreement did not require it.
Results
- Labeling accuracy above 90 percent for the full engagement
- Every dataset delivered on time against sprint cycles of 90 to 120 days
- All staff moved into the joint venture with no operational or billing disruption
- Full workforce continuity while the program’s leadership and structure changed
- Strong informal feedback on quality, program management, and flexibility