Patient data, transcripts, notes, intake forms, voice memos, is some of the most heavily regulated information in the US. Every call to a cloud API is a potential BAA question, a new breach surface, and an audit trail on someone else's server. A model that processes data locally eliminates the third-party processor entirely. That doesn't automatically satisfy HIPAA, but it dramatically simplifies the compliance conversation.
Running a local model doesn’t make you HIPAA-compliant by itself. What it does is remove one of the more complicated variables from the equation: the third-party processor.
When patient data leaves your building to reach a cloud API, you take on new obligations: a BAA with the vendor, reliance on their security posture, their breach notification timeline. Large health systems have compliance teams for this. A solo physician doesn’t.
The practices that benefit most are doing high-volume, structured tasks: drafting notes from voice, summarizing intake packets, generating template-based patient education materials. Formatting is what the model handles. The clinician handles the judgment.
My medical experiments focus on tasks with high documentation volume, clear quality standards, and a human-in-the-loop review step. SOAP notes are the natural starting point: the format is well-defined, a bad first draft just gets edited, and every note you don’t retype by hand is time back at the end of the day.
Future experiments will cover intake summarization, referral drafting, and the practical tradeoffs between model size and output quality for clinical text.
Every experiment on this site is a technical test, not a compliance recommendation. The results tell you whether local hardware can do the job at sufficient quality and speed. Whether any of these setups belongs in your specific practice, given your state, your specialty, your malpractice coverage, and your EHR system, is a question for your compliance team and legal counsel.
Read the Scope & Disclaimers page before making any decisions based on this content.
A base M4 Mac mini running a quantized Llama 3.1 8B model can draft structured SOAP notes from rough voice transcripts at a speed and quality that makes editing faster than writing from scratch, for most visit types.
The experiments here cover what's technically possible with local hardware. Your specific regulatory obligations, your state's rules, your specialty's requirements, your malpractice carrier's stance, your EHR vendor's terms, are questions this site cannot answer.
See the Scope & Disclaimers page.