Healthcare

Clinical notes, lab results, referral letters, and insurance paperwork are full of protected health information. Sending them to a cloud AI service means a third party processes that data, which brings agreements, reviews, and risk with it. Smartloop runs its models on your own machine, so you can use AI on patient documents while they stay on the device.

This page shows how to set up Smartloop for healthcare work and the tasks it is best suited to.

Why private AI in healthcare

Concern	With Smartloop
Patient data stays on the device	Documents, images, prompts, and answers are processed locally. Nothing is sent to a hosted model.
Fewer third parties	There is no cloud model provider handling the data, so no additional processor to vet for inference.
Retention and training	Patient data is not used to train anyone else's model, and there is no provider-side log of prompts.
Traceable answers	Every answer lists the documents it drew from in a REFERENCES block, so a summary can be checked against the record.
Works on ordinary hardware	Models are sized for everyday laptops and desktops — no GPU server or data-center contract needed. See Local AI.

Keep patient projects fully offline

Two features reach off your machine, and only when you use them: web search sends the search query, and connections send what a tool call needs to the service you connected. For any project holding patient data, turn search off in Project Settings and leave connections off. Everything else stays local.

Running locally keeps data out of third-party hands, but it is one part of compliance, not all of it. Device encryption, access controls, and your organization's policies still apply.

Try it with the starter project

The quickest way to see this is the healthcare starter, Lakeview Primary Care: a project for a fictional practice holding the records of one of its patients, with two skills for working through them.

Download Lakeview Primary Care (zip, 260 KB)

1. In Studio, click + New Project.
2. Choose Import, then pick the zip (or drop it onto the dialog), and click Create. The project is named Lakeview Primary Care.
3. Turn web search off with the toggle under the prompt (⌘⌥S, or Ctrl+Alt+S), so the project works only from the patient's records.

The documents are indexed in the background; they show as INDEXING in the Documents panel for a few seconds.

Document	What it is
Discharge-Summary-2026-01-18.pdf	First heart failure admission, with the discharge medications
Discharge-Summary-2026-06-09.pdf	Second admission, where several medications changed
Lab-Results-2025-2026.pdf	Cumulative lab report: HbA1c, eGFR, NT-proBNP, electrolytes
Referral-to-Nephrology-2026-09-30.docx	Referral letter from the primary care physician
New-Patient-Intake-Form-scan.jpg	A handwritten intake form, scanned — read by the vision model

Then try:

• "What changed in the medication list between the January and June discharge summaries?"
• "What allergies are documented, and what reaction did each cause?"
• /pre-visit-summary to prepare a one-page summary for the next visit.
• /extract-intake on the scanned intake form.

The practice, patient, clinicians, and facilities in the starter are made up; no real patient data is included.

Organizing projects

Each project has its own documents, memory, and agent process, isolated from every other project. Use that to keep data separated the way your organization requires:

• Per patient or case — for a care coordinator or a practice reviewing an individual's history.
• Per workflow — for example, a prior-auth project for payer correspondence, or a guidelines project holding clinical protocols with no patient data at all.

Add files through the Documents panel or by copying them into the project folder (⚙️ → folder icon). PDFs, Word, Excel, and images are supported — see Bring your documents.

What you can do

Summarize a patient's history

With a patient's records in a project, ask:

• "Summarize this patient's history: active problems, current medications, allergies, and recent procedures."
• "List every HbA1c result with its date, oldest first."
• "What changed in the medication list between the January and June discharge summaries?"

Each answer cites the documents it came from, so every value can be traced back to the record.

Read scans, forms, and faxes

Much of healthcare paperwork arrives as scans, faxes, or photos. Attach one to a prompt and the vision model reads it on your device:

• "Extract the patient name, date of birth, insurer, and member ID from this intake form."
• "What tests does this referral request, and who is the referring provider?"
• "Read this lab report and list any results flagged out of range."

Prepare letters and paperwork

Ask for drafts grounded in the record instead of written from scratch:

• "Draft a referral letter to cardiology summarizing the relevant history and the reason for referral."
• "Write a prior authorization request for this MRI, citing the documented failed conservative treatment."
• "Rewrite the discharge instructions in plain language at a sixth-grade reading level."

Give a thumbs up to a letter or summary in the style you want and it is kept in the project's memory, so later drafts follow it.

Look up your own protocols

Keep clinical guidelines, formularies, and internal procedures in a project of their own and ask questions against them — "What is our protocol for a needlestick injury?" — with the answer pointing to the page it came from.

A visit summary skill

Repeated documentation tasks are a good fit for a skill. This is the pre-visit-summary skill from the starter project; it turns a patient's records into a summary in a fixed format:

---
name: pre-visit-summary
description: When the user asks to prepare for a visit or summarize the patient, produce a one-page pre-visit summary from the patient's records in this project
---

# Pre-visit summary
Prepare a one-page summary of the patient's records for the clinician's next visit.

## Sections
- Active problems
- Current medications and doses, using the most recent discharge summary
- Medication changes since the previous admission, and why
- Allergies and reactions
- Results trend: HbA1c, eGFR, NT-proBNP and potassium with dates, marking any out of range
- Open referrals and outstanding follow-ups

## Guidelines
- Use only information found in the project's documents, and cite the document for each item
- Write "Not documented" when a section has no information; never infer
- Keep it to one page
- Do not suggest a diagnosis or treatment

Run it from the prompt with /pre-visit-summary, and adjust the sections and the results it tracks to your specialty.

Automating with the API

The local agent serves an OpenAI-compatible API on localhost:38540, so the same extraction can run over a batch of documents from a script — pulling structured fields out of a folder of intake forms, for example. Requests stay on the machine:

from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:38540/v1", api_key="not-needed")

response = client.chat.completions.create(
    model="sl-mini",
    messages=[
        {"role": "user", "content": "List every medication in these records with dose and start date, as a table."}
    ],
    extra_body={"project_id": "<your-project-id>"},
)

print(response.choices[0].message.content)

See Local AI for finding a project's ID, multi-turn requests, and reading citations from the response.

info

Smartloop is a documentation and information tool, not a medical device, and does not provide diagnoses or treatment advice. A qualified clinician should review every answer against its cited sources before it is used in care.