---
title: "Private Clinical AI — Whitepaper"
description: "Clinical documentation, summaries and literature research with AI, without patient data leaving the machine. Evidence-grounded, and not for diagnosis."
url: "https://royalsoftworks.com/resources/healthcare-ai-on-device/"
source: "https://royalsoftworks.com"
format: "markdown"
note: "Markdown rendering of the HTML page at `url`. Same content, same canonical URL."
---

[All resources](https://royalsoftworks.com/resources/)

# Private Clinical AI

## Evidence-Grounded, On-Device Clinical Documentation

**A Royal SoftWorks whitepaper · AssistantGeneral**

---

## Abstract

Healthcare imposes two requirements that cloud AI struggles to meet at once: protected health information must not leave the institutional boundary, and anything written for the record must be tied to real evidence rather than plausible-sounding invention. AssistantGeneral keeps PHI on the clinician's device — including on-device dictation — and grounds research answers in primary literature with an explicit check against retracted studies. This paper describes the architecture, states the intended purpose plainly — documentation, summarisation, research and administrative support — and is candid about where human judgment remains mandatory.

---

## 1\. Two hard requirements

**Privacy.** PHI is governed by HIPAA, GDPR, and institutional policy. Transmitting it to a commercial AI cloud requires a BAA and a residency story, and many institutions forbid it regardless. Most leading tools send the record to a vendor cloud; practices that cannot allow that need an on-device option.

**Evidence.** Clinicians will not — and should not — trust an assistant that invents citations or relies on studies that have since been retracted. An ungrounded statement in a clinical document is a patient-safety issue.

## 2\. Privacy by architecture

AssistantGeneral runs the whole stack on the device:

- The **model**, the **document index**, and the **embeddings** are local.
- **Dictation is on-device** — Whisper transcribes the spoken note locally, so the audio and transcript never leave the machine.
- **Document generation** (letters, summaries) runs in a local sandbox.
- A cloud model is attached only if institutional policy permits it.

For a practice or department, the governance plane — policy, audit and shared knowledge bases, with company sign-in in pilots — runs on an admin console on one of the practice's own machines. PHI stays inside the institutional perimeter end to end.

## 3\. Evidence by design

The **Doctor agent** is restricted to research-grade sources and carries a mandatory disclaimer ("general information, not medical advice or diagnosis; seek urgent care for emergencies"). Its research tools are built for rigor:

- **`pubmed_search`** and **`europepmc_search`** — peer-reviewed literature, cited by PMID.
- **`clinical_trials_search`** — registered trials, cited by NCT identifier.
- **`retraction_check`** — run _before_ relying on a study, so the agent never cites withdrawn or corrected evidence.

The agent is instructed to prefer systematic reviews, meta-analyses, and major-body guidance, and to cite specifically.

## 4\. The groundedness gate

On top of source selection, AssistantGeneral applies a **groundedness gate** to its answers. After the model drafts a response, the gate checks each factual claim against the evidence actually gathered. When several claims are unsupported, it **re-retrieves targeted evidence** and sends the draft back to be supported, qualified, or removed — a bounded, closed re-RAG loop; a single stray claim is flagged so the answer hedges it. In a clinical setting this is the difference between a summary that _sounds_ authoritative and one that is tied to the record and the literature.

## 5\. Capabilities, mapped to clinical paperwork

The workflows this paper covers are documentation, research and administration:

- **Clinical Note Summariser** — structured summary of a clinical note.
- **Referral Letter Drafter** and **Discharge Summary Generator** — documents drafted from structured inputs.
- **Medical Research Summariser** — literature synthesis from the sources above.
- **Clinical Guidelines Lookup** — over the guideline documents you load into the knowledge base.
- **ICD-10 Code Mapper** — documented conditions mapped to ICD-10 codes for the record.
- **Consent Form Explainer** — plain-language explanation of a consent form.

When a question concerns the patient's own record, the agent first gathers it from the local knowledge base before writing — keeping the document grounded in the actual chart.

## 6\. The interface matches the setting

Two clinical workspaces, because a ward and a single practice do different paperwork. The **Hospital** workspace is organised around an _episode_ of care — clinical-note summaries, discharge summaries and coding for the record, producing a discharge summary. The **Clinic** workspace is organised around a _consultation_ — note summaries, referral letters and plain-language consent explanations, producing a referral letter. Each limits its chats to the collections that belong to that episode or consultation, enforced in retrieval, and carries a notice that cannot be dismissed: a clinician decides. Practice packs add the local reference material for the jurisdiction the practice works in.

## 7\. Intended purpose — stated plainly

AssistantGeneral's clinical use is **documentation, summarisation, research and administrative support**, always reviewed by a qualified professional. It is **not intended for diagnosis, triage, treatment decisions or clinical decision support**, and it is not a medical device. Presets for those purposes that remain in the application's general library fall outside this intended purpose and must not be used for clinical decisions. Every specialist interaction begins with a disclaimer.

On-device dictation uses a general-purpose speech model, not one trained on medical vocabulary: check every transcript before it goes into the record.

The combination of on-device PHI handling, evidence-grounded sourcing, retraction checking, and the groundedness gate is what makes the tool usable for paperwork in a setting where both privacy and safety are non-negotiable.

---

_Evaluate it with the network disabled: dictate a note, summarise it, and ask a research question — and watch the retraction check filter a withdrawn study. Contact Royal Softworks._

## Evaluate AssistantGeneral

See the product this paper describes — a private, local-first AI agent with an enterprise control plane that runs on your own infrastructure.

[Explore AssistantGeneral](https://royalsoftworks.com/products/assistant-general/) [Talk to us](https://royalsoftworks.com/contact/)

<!-- verse:royalsoftworks
     Joil 2:12
     Zato još govori Gospod:Obratite se k meni svim srcem svojim i posteći i plačući i tužeći.
     https://www.wordproject.org/bibles/verses/serbian/02_prayer.htm
-->
