Grae SI

Sandbox profile

Utilization-review packet drafts

High intensity

A health system drafts medical-necessity packets inside its own VPC. A physician still decides. The model does not submit the packet.

Vertical
Healthcare providers
Scale
Enterprise
Organization
Lumen Health
Shape
Regional system, 8 hospitals, a central utilization-review team.

The business problem

Utilization review nurses assemble records, criteria, and a narrative for procedures that need a payer conversation. The work is slow, the records are PHI, and a public model is not an acceptable processor. The system wants a draft packet a physician reviews. It does not want the model to approve or deny care.

What Grae SI recommends

Family

Qwen 2.5 Instruct or Llama 3.1 Instruct

Size

7B–8B with LoRA. One 24–48 GB GPU, or two smaller GPUs in the private VPC.

Why this size

Packets are longer than a clinic note and cite criteria. A 3B model drops conditions. A 70B model adds a cluster the system does not need for a draft that a physician rewrites. 7B–8B is the smallest size that holds a multi-page record and still fits one private node.

License

Llama community terms restrict some uses. Qwen’s license must be read the same way. The card flags restrictions before training starts.

Data controls

PHI

Charts, labs, and imaging reports are PHI. Training uses a de-identified sample of closed reviews. Live inference stays in the VPC and is PHI.

De-identification

Expert determination or safe harbor, chosen by HIM, for the training set only. Dates of service can be shifted. Live packets keep real dates because the physician needs them.

Access

UR nurses and the reviewing physician. Break-glass is the EHR’s, not a new one. Model logs are restricted to security and HIM.

Residency

The VPC is in the system’s US region. No cross-region replica of prompts. Weights do not sit in a vendor account.

Retention

Prompt logs follow the UR record schedule. Adapters are kept for the life of the released version plus the system’s model-retirement rule.

Risk classification

Intensity is a planning tier for evidence. It is not a certification and not a regulatory classification.

  • PHI plus a narrative that can change whether a procedure is presented as necessary.
  • A person reviews and signs, which keeps the model out of an automated denial, but the output is still high intensity.
  • If the draft were submitted to a payer without that review, the same tier would apply and the control would be gone.

Compliance requirements to prepare for

HIPAA

The system is a covered entity. Business associates, if any integrator is used, need a BAA. The model host should be inside the system so the model vendor is not a processor of PHI.

SOC 2 and ISO/IEC 27001

Expected by payers and by the system’s own audit plan. Grae SI helps gather the AI-specific evidence. The attestation is the system’s auditor.

ISO/IEC 42001 and NIST AI RMF

Intended use, impact, and a human in the loop are the spine of the record.

CMS and payer rules

Medical-necessity criteria and who may sign are payer and medical-staff rules. The model does not interpret coverage.

Grae SI helps assemble the evidence. It does not certify HIPAA, FDA, ISO, SOC 2, ICH, NAIC, GDPR, DPDP, or any other framework.

Validation KPIs

These are targets to lock before go-live. They are not measured outcomes from a customer deployment.

Clinician edit acceptance

80% of drafts accepted with moderate edits on 150 historical packets

UR prose is picky. The test is whether a physician keeps the facts, not whether the tone matches a template.

Unsupported assertion

Under 2% of sentences state a finding absent from the source packet

Hallucinated history is the failure mode that matters.

Criteria citation

Every cited criterion appears in the source criteria sheet

The model must not invent a guideline.

Human review

100% of outbound packets have a recorded physician sign-off

High tier. The gate is a person, and the report records the gap if that sign-off is missing.

Deployment architecture

DeploymentPrivate VPC, no public egress for prompts

EHR and the document store stay in the system’s network. A private endpoint serves the model. Egress for prompts and weights is blocked. Administrators patch the host from a management subnet.

  1. 01

    EHR and document store

    Records are read over the system network by a service account with UR scope.

  2. 02

    Prep zone

    De-identification for training runs in the same VPC. Outputs are a dataset, not a ticket.

  3. 03

    Training

    Grae SI job on one private GPU node, or the system’s existing cluster via the customer-GPU adapter.

  4. 04

    Serving

    vLLM on a private endpoint. Security group allows the UR app only.

  5. 05

    UR workbench

    Draft appears as an unsigned packet. Submission stays a human action in the existing tool.

Each box is a component in the customer’s network. Arrows are the order of the path, read left to right on a wide screen and top to bottom on a phone.

Where this sits in the stack

The same ten layers appear on every profile. The mark says whether Grae SI provides the layer, shares it, or leaves it with you. Person-weeks are the buyer’s effort after the runtime is deployed, not a project plan for the whole company.

  • Grae SI providesGrae SI provides the layer
  • SharedShared. Grae SI supports part of it.
  • You ownYou own the layer after install

  1. Layer 1

    Infrastructure and compute

    You own

    Security architecture review for a private GPU node in the VPC the system already runs.

    Buyer effort after the runtime is up: 2–4 person-weeks

  2. Layer 2

    Data ingestion and prep

    Shared

    Grae SI scores a sample extract. The EHR feed, and what may leave the chart, is yours.

    Buyer effort after the runtime is up: 2–5 person-weeks

  3. Layer 3

    Data controls and governance

    Shared

    Masking preview for names, MRNs, and dates. Access, BAA scope, and retention stay with the covered entity.

    Buyer effort after the runtime is up: 1–3 person-weeks

  4. Layer 4

    Model selection and training

    Grae SI provides

    7B–8B instruct model, LoRA, on the system’s cluster or a single private node. Same Grae SI job spec either way.

    Buyer effort after the runtime is up: 0.5–1.5 person-weeks

  5. Layer 5

    Evaluation and validation

    Shared

    High-tier suite, a 150-packet gold set, and a recorded human sign-off before release.

    Buyer effort after the runtime is up: 2–4 person-weeks

  6. Layer 6

    Agents and orchestration

    Shared

    Compliance agents that classify the use case and name the frameworks to prepare for. Any production agent that calls tools inside your applications.

    Buyer effort after the runtime is up: 1–2 person-weeks

  7. Layer 7

    Deployment and serving

    Grae SI provides

    Safetensors, GGUF, a model card, and an offline bundle for Ollama, vLLM, or llama.cpp. Moving that bundle into the zone and keeping the host patched.

    Buyer effort after the runtime is up: 0.5–1.5 person-weeks

  8. Layer 8

    Monitoring and drift

    Shared

    The system’s SOC watches the endpoint. Grae SI’s registry is the version lineage, not the SIEM.

    Buyer effort after the runtime is up: 2–4 person-weeks

  9. Layer 9

    Application and integration

    You own

    Integration with the UR workbench is the long pole. The model service itself is the smaller piece.

    Buyer effort after the runtime is up: 6–10 person-weeks

  10. Layer 10

    Organizational validation and sign-off

    You own

    Privacy, medical staff, and compliance each sign the intended use. Grae SI does not convene that committee.

    Buyer effort after the runtime is up: 3–5 person-weeks

Buyer effort for this profile

14–22 person-weeks after the VPC and GPU quota exist

The model run is days. The gold set, the workbench hook, and the sign-off are the schedule.

  • Clinical informatics
  • HIM or privacy
  • Security engineer
  • UR medical director
  • Integration engineer

Quantitative benefits

Illustrative estimate for this sandbox. It is not a measured customer result. Assumptions for this profile are listed under the figures.

Time to a reviewable package

8–12 weeks

Assumes GPU quota and an EHR interface analyst are already on staff. A greenfield private-model program at this size is often two to three quarters.

Cost versus a cloud API

About $18k–$40k a month avoided

Illustrative: 20,000 packets a month, long context, blended hosted price $0.9–$2.00 a packet. Residency rules would block that API anyway.

Compute

$1.5k–$3.5k a month

One private GPU node reserved in the system’s cloud account, or depreciation on hardware they already bought. Not a Grae SI fee.

Validation writing

About 3–5 person-weeks saved

Model card, tier rationale, and eval export exist. Payer-facing policy language is still written by the system.

Accuracy target

80% moderate-edit acceptance, under 2% unsupported sentences

Targets for the gold set before go-live.

Assumptions

  • Illustrative estimate for this sandbox. It is not a measured customer result.
  • Volume is 20,000 review packets a month across 8 hospitals. A single service line is a fraction of the API comparison.
  • The model never submits a packet and never recommends a denial code.
  • Hosted-API prices move. The comparison is an order-of-magnitude check, not a quote.

All sandbox profiles · The full stack