PRIVATE LLM DEPLOYMENT FOR HEALTHCARE

HIPAA LLM and RAG Development That Keeps PHI Under Your Control

NonStop builds LLM and RAG systems for health systems, health tech companies, diagnostic labs, genomics labs and life sciences teams. Run them on a BAA-covered model service, in your own cloud or fully on-premise.

90+ clients since 2015
On-premise, self-hosted or cloud
Anthropic, OpenAI, Gemini and open-weight models
Engineered for HIPAA, SOC 2, GDPR, and the CLIA and CAP environment

A BAA Covers the Model. PHI Lives in Six More Places.

In a healthcare RAG system, patient data touches seven layers. NonStop builds a control into each one.

1
PROMPT

Identifiers the task does not need are stripped before the model call.

2
IDENTITY

Every query is tied to a named user and role.

3
RETRIEVAL

Search returns only what that user is allowed to see.

4
MODEL

Runs under a BAA or inside your own environment.

5
ANSWER

Every claim cites its source, with human review where risk calls for it.

6
LOGS

PHI is redacted before anything is logged.

7
TEST DATA

Cases are de-identified before they enter any dataset.

Embeddings count too. Researchers recovered 92% of short text inputs exactly from their embeddings (Morris et al., EMNLP 2023).

What Healthcare Teams Get From a Private LLM Build

Answers Clinicians Can Verify

Every answer links to the guideline, SOP or prior case it came from.

Search That Respects Permissions

Permission-aware retrieval mirrors your source systems, so nobody sees records they should not.

PHI Stays Where You Choose

Cloud model service, vendor API, your own cloud account or on-premise.

Evidence for Your Privacy Officer

Audit trails, redacted logs and documented controls, mapped to 45 CFR §164.312.

Results From NonStop Healthcare AI Builds

4x
faster prior authorization decisions on a PHI-safe LLM platform with a FHIR-native decision engine
Under 2 min
per requisition form with on-premise AI intake (SmartReq), down from 15 to 25 minutes
3x
faster guideline-based evidence retrieval from RAG-powered clinical knowledge systems
80%
time-to-report saved with AI-assisted variant interpretation, 100% human-in-the-loop preserved

Where to Run Your LLM With PHI

On-Premise

Open-weight model and vector search on your own hardware.

Self-Hosted in Your Cloud

Open-weight model in your own AWS, Azure or Google Cloud account.

Cloud Model Service

Amazon Bedrock, Azure OpenAI or Google Cloud Agent Platform (Vertex AI), under the provider's BAA.

Model Vendor API

Anthropic's Claude API or OpenAI's API, under the vendor's BAA for eligible features.

Every path still needs a risk analysis and Security Rule controls. BAA scope varies by service and feature, so NonStop confirms it before any PHI moves.

How a HIPAA LLM Project Works With NonStop

1

AI Architecture Review

45 minutes. Bring one question you want an LLM to answer. A senior engineer maps where PHI would flow.

2

PHI Boundary and Plan

Hosting path, retrieval design, evaluation approach and the controls each layer needs, in a phased plan.

3

Build, Test, Hand Over

Engineers embed on-site in the US or work remotely, test on real workflows, and hand over the system with full IP ownership.

Healthcare RAG Use Cases by Team

Genomics Labs

Search your own variant interpretation history, as in GENVAR, deployed on-premises.

Diagnostic Labs

Read requisitions and answer SOP questions without patient data leaving the lab.

Health Systems and Health Tech

Guideline and payer policy search, and prior authorization support with clinician sign-off.

Pharma and Biotech

Search protocols, SOPs, regulatory documents and PubMed with source citations.

HIPAA LLM Readiness Checklist

Score any LLM build against it, including ours.

  • A signed BAA covers the model service, and the features or endpoints in use are in scope.
  • Every other vendor that stores PHI, such as the vector database, has a BAA or runs inside your environment.
  • Embeddings made from PHI are encrypted and access-controlled.
  • Retrieval enforces each user's permissions from the source system.
  • Separate indexes keep tenants or client labs apart.
  • PHI detection runs on inputs and outputs.
Show all 12
  • Logs redact PHI and follow a set retention period.
  • Every answer cites its source.
  • Test and training data are de-identified under 45 CFR §164.514(b).
  • Red-team tests cover prompt injection, including instructions hidden in retrieved documents.
  • Model versions are pinned, and changes are tested against agreed acceptance criteria.
  • A named person reviews output wherever clinical or quality risk calls for it.

Common Questions About HIPAA and LLMs

Is there a HIPAA compliant LLM?
No. HIPAA applies to organizations, not software, and HHS does not recognize product certifications under the Security Rule. An LLM system can support compliance when the model runs under a BAA or inside your own environment. Risk analysis, access control, audit controls and encryption must then cover every layer that touches PHI: prompts, embeddings, vector stores, logs and test data.
Can I use the OpenAI API with PHI?
Yes, under a BAA with OpenAI and only on eligible endpoints. OpenAI lists its API with Modified Retention as HIPAA-eligible. Consumer ChatGPT plans are not covered.
Are vector embeddings considered PHI?
HIPAA does not name embeddings, so treat them as PHI whenever they come from PHI. Research at EMNLP 2023 recovered 92% of short text inputs exactly from their embeddings.
Does a vector database vendor need to sign a BAA?
Yes, if it stores or processes PHI on your behalf, because it meets the definition of a business associate under 45 CFR §160.103. A self-hosted vector store inside your environment removes that extra vendor.
RAG or fine-tuning for clinical data?
Start with RAG. It keeps PHI in stores you control, applies permissions per user and cites its sources. PHI inside fine-tuned model weights cannot be selectively deleted, so fine-tune only on de-identified data.
Is a self-hosted LLM automatically HIPAA compliant?
No. Self-hosting removes the model vendor from the data flow, but the Security Rule still applies to wherever the model runs.

Bring One Question You Want an LLM to Answer

A NonStop senior engineer traces it through all seven layers, recommends a hosting path and names the controls still missing.

45 minutes. No sales pitch. No commitment required. Please do not include PHI in this form.

Not ready for a call? Take the 15-Minute AI Readiness Self-Assessment