Production ML where being wrong is expensive
Distributed ML and governed data systems across DEKA programs including Roxo, the FedEx SameDay Bot, and the FDA 510(k)-cleared twiist insulin delivery system.
- 80% less annotation work. Across my DEKA work.
- 35% lower end-to-end latency. Across my DEKA work.
- 60% better OCR root-cause accuracy. Across my DEKA work.
- 43% fewer production bugs. Across my DEKA work.
- 57% faster releases. Across my DEKA work.
Context & problem
DEKA Research & Development builds machines where being wrong is expensive. My work spanned two of its defining programs: Roxo, the FedEx SameDay Bot, an autonomous last-mile delivery robot developed by DEKA in collaboration with FedEx, and the twiist Automated Insulin Delivery system, developed by DEKA, commercialized by Sequel Med Tech, and FDA 510(k)-cleared.
Across both, ML teams were drowning in near-duplicate data: annotation effort scaled with raw volume, not information content.
My role
Led design and development of the image deduplication system; designed and implemented four distributed ML workflows from the ground up; built the HIPAA-compliant OCR root-cause analysis system; drove the CI/CD and TDD transformation.
Architecture
Embedding-based similarity over an AWS OpenSearch backbone for deduplication; Airflow-orchestrated distributed workflows in Python and C++ on Kubernetes. Working under HIPAA and medical-device constraints shaped these systems end to end. Governance was not a slide in the deck. It was the operating condition.
Measured result
Across my DEKA work: −80% image annotation workload, −35% end-to-end latency, +60% OCR root-cause accuracy, and a CI/CD transformation that reduced production bugs by 43% and made releases 57% faster.
Reach out to learn more.