A Gulf South health system outgrew COCO Annotator as its nasal-endoscopy AI program scaled. We rebuilt its catalog on FastAPI and Svelte, moved labeling to CVAT with bi-directional sync, and added a grading engine that scores labelers against expert benchmarks.
Discuss Your Labeling PipelineA nonprofit health system in the Gulf South running an internal endoscopy AI program
Catalog redesign, COCO to CVAT migration, bi-directional sync, labeler grading
FastAPI, Svelte, IBM Carbon, CVAT, PostgreSQL, Redis
Delivered in a five-week phased engagement
The hardest part was making COCO retirement safe: keeping the internal clinical data model intact while a different labeling platform took over underneath it, without losing a single patient's annotation history in the handoff.
Labelers coordinated per project with real-time progress tracking
Patient or annotation records lost in the COCO to CVAT migration
From audit and discovery to grading and hardening
Test coverage target on core catalog features
It handled bounding boxes well. It could not handle multi-attribute clinical workflows, dozens of parallel labelers, or a data model that kept changing as clinical needs evolved.
Projects ran 20 to 25 clinical labelers at once, with no reliable way to assign work by attribute or track "in progress" and "completed" states.
Teams could not filter by category, labeler, patient, or assignment state. Exact text search by code or category did not exist.
COCO's structure did not match the internal model of patients, encounters, and observations, and the gap widened with every new clinical requirement.
Data moved through manual spreadsheet uploads, and ML teams lost sight of dataset state exactly when the program needed to scale.
FastAPI, SQLModel, Celery, Redis, PostgreSQL and Python on the backend; Svelte, TypeScript and IBM Carbon Design System on the frontend; CVAT as the labeling platform.
The engagement ran in five stages: audit and discovery, standardization, push sync and migration, pull sync and advanced search, then grading and hardening. The first job was consolidating models that had drifted across the codebase.
Push and pull engines move only what changed, so the catalog and CVAT stay aligned without spreadsheet uploads. Legacy COCO datasets came across with their clinical attributes intact.
We extended the Svelte and IBM Carbon interface into dedicated modules, each with the same filtering sidebar, so a lead can find any assignment, file or labeler the same way on every page.
A grading engine scores novice labelers against expert benchmarks, and a test suite guards the sync and grading flows as new endoscopy modalities are added.
We migrate annotation platforms without breaking the clinical data model underneath them.
Partner with us to design, build, and scale digital solutions that drive better outcomes.
Global Tech Teams LLC, 525 Washington Blvd, Industrious at Newport Tower, Jersey City, NJ 07310, United States.
Let’s discuss your goals, workflows, and next steps in a focused consultation call.
Schedule a Call 
BLOGS 
NEWSROOM 
CASE STUDIES 
WEBINARS 
PODCASTS 
ASSET HUB 
EVENT CALENDAR 
