Under contract HT003826PE001, we built a multi-cloud data ecosystem for the Defense Health Agency and a proof of concept on it. Pipelines bring health records from separate systems into governed data products. AI agents analyze the records and answer questions across documents. In a portal, service members see their complete treatment record, flag errors, and certify it. It runs at a scale of one million records, and every value traces back to its source.
Before a service member leaves active duty, their military treatment record becomes the service treatment record that follows them into veteran status. That record draws on operational, treatment, and purchased-care sources held in separate systems.
The agency wanted one place where a service member could see the complete record, check it for gaps and errors, and certify it before transition.
We built the data ecosystem, the AI agents that work on it, and the portal service members use. We also delivered documentation for building the same approach across the agency’s data: future-state architecture, security posture, operating model, technical specifications, and a report of what we learned.
The proof of concept showed four things a service member can do:
Source data lands unchanged in raw storage. A curation step turns it into governed data products in a common clinical schema and records the lineage of every value. A composition step assembles each service member’s record and applies rules for which source takes precedence. AI agents extract structured values from documents and notes, resolve conflicts between sources during assembly, and answer questions across documents, including whether something happened at all.
The system is multi-cloud and defined entirely as code, so it can be rebuilt in an empty cloud account from one pipeline run. It runs at a scale of one million records. A data catalog and a lineage view let follow-on teams find the data and trace any value back to its source.
A service member sees only their own record, enforced at the row level, and each AI service receives only the fields it needs. Records in sensitive categories, such as substance use and HIV, are flagged when they arrive and shown only after an explicit action. Every disclosure is logged for HIPAA accounting, and the question-answering service answers only about the requesting member’s record, with every prompt and response logged.
Controls were designed against HIPAA, with NIST guidance and DoD overlays, and include encryption, least-privilege access, and network isolation. The proof of concept used synthetic health records, with controls built to the standard a production system needs.
Engineering started on the first day, alongside design. At week 7 we demonstrated a basic version that ran through every part of the system; the demonstrations at weeks 10, 12, and 14 added scale, quality, and the hardest questions. Documentation drafts arrived at week 10 so agency reviewers had time to comment before the final evaluation.
The service member reviews the record and certifies it. Flagged inaccuracies go to people to resolve.
The Defense Health Agency is available as a reference. Ask Linda Apsley for details.
Start with a 30-minute conversation with Linda about your records, reviewers, and security requirements.
Talk with us