Sandbox profile
After-visit note drafts
Medium intensityA small clinic drafts after-visit summaries on a machine it owns, without sending the chart to a public model API.
- Vertical
- Healthcare providers
- Scale
- Small company
- Organization
- Harbor Street Clinic
- Shape
- Independent primary care, 12 clinicians, one billing office.
The business problem
Clinicians dictate or type a short visit note, then spend the evening turning it into language a patient can read and a coder can bill from. A hosted chatbot would be faster and would also put names, medications, and problems on someone else’s infrastructure. The clinic wants a draft in the chart, reviewed by the clinician who saw the patient, and a bill for hardware they can explain to a partner meeting.
What Grae SI recommends
Family
Qwen 2.5 Instruct, or Phi-4 mini if the host is a single workstation
Size
1.5B–3B for a clinic workstation. Step up to 7B only if a 24 GB GPU is already on site.
Why this size
The task is a constrained rewrite of a note the clinician just wrote, not open-ended diagnosis. A small instruct model with a LoRA on de-identified visit pairs is enough, and it fits a machine the clinic can buy without a platform team. A 70B model would raise cost and would not remove the clinician review.
License
Confirm the upstream license before download. A permissive or research-restricted license is printed on the model card. Grae SI does not grant it.
Data controls
PHI
Visit notes are PHI. Training rows are de-identified: names, MRNs, phone numbers, and exact dates stripped or shifted. The live prompt still contains PHI and never leaves the clinic network.
De-identification
Safe-harbor style removal for the training extract. The clinic’s privacy lead confirms the method. Grae SI shows a redaction preview on a sample, it does not certify the extract.
Access
Only the clinician of record and the privacy lead can open a draft. Service account for the model is local. No shared login.
Residency
Weights, prompts, and logs stay on the clinic server. No region failover to a public cloud.
Retention
Prompts follow the medical-record schedule the clinic already uses. Training extracts are deleted after the adapter is accepted, on a written date.
Risk classification
Intensity is a planning tier for evidence. It is not a certification and not a regulatory classification.
- The note contains PHI, so this cannot sit at low intensity.
- The model drafts. The clinician signs the note. It does not diagnose, prescribe, or send a result to a patient on its own.
- If a later version filed the note without a review, the tier would move to high.
Compliance requirements to prepare for
HIPAA
Privacy, Security, and Breach Notification apply if the clinic is a covered entity. Minimum necessary, access logs, and a de-identification method belong in the prep record.
SOC 2
Useful if a health-system partner asks. Grae SI can collect evidence of who trained what. It does not issue a SOC 2 report.
ISO/IEC 27001 and 42001
27001 for the clinic’s security controls. 42001 for an AI management system: intended use, risk record, and how data, training, and evaluation connect.
State privacy
If the clinic treats patients in a state with a consumer health privacy law, the notice and the retention schedule still come from counsel.
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.
Draft acceptance
At least 85% of drafts signed with light edits on a 200-note gold set
The clinician’s edit distance is the practical test. A fluent wrong medication is a failure even if the prose scores well.
Medication fidelity
Zero invented medications on the gold set
The draft may only use drugs present in the source note.
Identifier leak in training extract
0 names or MRNs in a manual review of 50 training rows
The live prompt may contain PHI. The training file should not.
Review time
Median clinician review under 90 seconds
If review takes as long as writing the note, the workflow has no point.
Deployment architecture
A single server on the clinic network serves the model. The EHR connector is a small service the clinic’s vendor or a contractor runs. There is no path from the note to the public internet.
01
EHR
Source note stays in the chart. A local connector copies the text to the model host.
02
Prep
Training extracts are masked on the same host. The live prompt is not masked, because the clinician must see the real note.
03
Grae SI runtime
Ollama or llama.cpp with the LoRA. Listens only on the clinic LAN.
04
Chart draft
The draft returns to an unsigned section. The clinician signs in the EHR.
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
- You own
Layer 1
Infrastructure and compute
One GPU server in the clinic or hospital network. Grae SI does not host the model.
Buyer effort after the runtime is up: 1–3 person-weeks
- Shared
Layer 2
Data ingestion and prep
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
- Shared
Layer 3
Data controls and governance
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
- Grae SI provides
Layer 4
Model selection and training
Qwen 2.5 1.5B–3B, or Phi-4 mini, through the Grae SI catalog and a local PEFT or Unsloth job.
Buyer effort after the runtime is up: 0.5–1.5 person-weeks
- Shared
Layer 5
Evaluation and validation
Medium-tier suite plus a 200-note gold set the clinic labels. A clinician samples the report before go-live.
Buyer effort after the runtime is up: 1–3 person-weeks
- Shared
Layer 6
Agents and orchestration
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
- Grae SI provides
Layer 7
Deployment and serving
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
- Shared
Layer 8
Monitoring and drift
Registry holds the version. The clinic watches edit rate monthly. There is no central Grae SI telemetry.
Buyer effort after the runtime is up: 0.5–1 person-weeks
- You own
Layer 9
Application and integration
One EHR side door, or even a paste-in screen at first. A twelve-clinician site should not start with a full interface engine.
Buyer effort after the runtime is up: 2–4 person-weeks
- You own
Layer 10
Organizational validation and sign-off
A privacy and clinical owner sign the intended use. Grae SI does not decide that the workflow is HIPAA-ready.
Buyer effort after the runtime is up: 1–4 person-weeks
Buyer effort for this profile
4–7 person-weeks after the server is racked
Most of the time is the gold set and the EHR hook, not the training run. A weekend GPU job is the small part.
- Clinic privacy lead
- One clinician reviewer
- EHR contractor or vendor analyst
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
3–5 weeks
Assumes the EHR vendor will open a read of the unsigned note. In-house, a first private model often takes a small clinic 4–6 months because nobody owns eval or licensing.
Cost versus a cloud API
About $1.2k–$2.5k a month avoided
Illustrative: 8,000 visits a month, two model calls each, blended hosted price $0.08–$0.15. Hosted chat also fails the residency rule this clinic set.
Compute
$40–$90 a month in power, or ~$0 if a workstation GPU is already on
One 16–24 GB consumer or workstation GPU. No reserved cloud instance.
Validation writing
About 1–2 person-weeks saved
The report, model card, and data checklist are generated. Counsel and the privacy lead still write the policy language.
Accuracy target
85% light-edit acceptance, zero invented drugs
A target for the gold set, not a result from a live clinic.
Assumptions
- Illustrative estimate for this sandbox. It is not a measured customer result.
- Visit volume is 8,000 a month. A quieter panel moves the API comparison down, not the ownership argument.
- The clinician still signs every note. No autonomous filing.
- De-identification of the training extract is reviewed by the clinic. The sample redaction in Grae SI is a preview.