Federal regulatory filings, generated on the client's own hardware
Mistral and Llama served inside a US healthcare provider's network behind an ASP.NET Core backend. Patient data never left their infrastructure.
Client work under NDA, so this describes the architecture rather than the client.
Findings in. Filings out.
A US healthcare provider was writing federal regulatory narratives by hand after every inspection. The documents go on file with a federal agency, so structure, citation format and tense are prescribed — a near miss is a failed filing.
Sending patient data to a hosted API was never an option.
So the models run on their hardware. Open-weight LLMs sized to the GPU they already owned, served behind an ASP.NET Core backend. Findings go in as structured entries, and every sentence in the output traces back to one of them — the system won't write a finding that wasn't observed.
What comes out is a draft. A reviewer edits it and files it. That boundary was fixed before any code was written.
Working on something in this territory? Tell us what you’re building