Techunison-style research blueprint

Frappe Health AI Scribe, Coding, and Billing Automation

Detailed research and implementation document for building an ambient clinical documentation platform on top of Frappe Health and ERPNext. This document covers architecture, source-code patterns, open-source research, security guardrails, coding extraction, and invoice automation.

Frappe Health

Core system of record

AI SOAP Notes

Primary use case

ERPNext Sales Invoice

Billing target

Open + Hybrid AI

Delivery model

Executive Summary

Research Findings

Model Strategy

Architecture

Implementation

Source Code

Security

Roadmap

Sources

1. Executive Summary

What should be built, why it matters, and where Frappe Health fits.

Recommended platform direction

Build a custom Frappe Health extension for AI-assisted clinical documentation rather than treating the AI scribe as an isolated tool. Frappe Health should remain the system of record for patient, encounter, medication, clinical coding, and billing workflows. The AI stack should sit beside it as an assistive service layer.

Best-fit approach: Browser or mobile audio capture → speech-to-text → structured SOAP JSON → draft Patient Encounter → optional ICD/CPT extraction → draft Sales Invoice → human review and submission.

Why this approach wins

Preserves full workflow control inside Frappe Health and ERPNext.

Supports both cloud and self-hosted AI models.

Allows phased rollout: notes first, then coding, then billing automation.

Fits Techunison’s enterprise healthcare implementation model.

Keeps the doctor in control by saving AI output as draft before submission.

Key design decisions

Decision Area	Recommended Choice
Clinical record ownership	Frappe Health Patient Encounter
Audio capture	Browser recorder / mobile app / room mic station
Speech engine	Whisper or equivalent STT layer
SOAP structuring	Prompted LLM returning strict JSON
Coding extraction	Separate coding prompt and validator
Billing logic	ERPNext item mapping with invoice draft
Safety mode	Mandatory clinician review before final submission

Do not let the model directly submit clinical records or invoices without validation, access control, and approval workflow.

2. Research Findings

Condensed from the provided research notes and validated against official Frappe/Frappe Health documentation.

Finding 1

Frappe already exposes REST APIs for DocTypes

That means Patient Encounter, Sales Invoice, Item lookups, and custom DocTypes can all be created or queried using standard REST endpoints without building a separate backend first.

Finding 2

Patient Encounter is the right clinical anchor

Encounter is where visit-level documentation, diagnosis coding, medication, and downstream billing context naturally belong.

Finding 3

Open source is viable, but hybrid is stronger

Open-source stacks such as scribeHC and local STT can reduce lock-in, but production-grade performance often improves when combined with stronger hosted reasoning models or managed inference infrastructure.

What the research strongly suggests

Area	Research Direction	Implementation Reading
Ambient documentation	Use a healthcare-specific capture flow, then normalize into structured SOAP JSON.	Keep transcript, note draft, and clinician edit history separate.
Open-source tools	Use scribeHC-style tooling for the listening layer and Frappe for system workflows.	Best when you want customization and self-hosting control.
Medical coding	Prompt the model for diagnosis and procedure suggestions in strict JSON	Always validate against your own Medical Code / Item master before billing.
Billing automation	Generate invoice drafts, not final submitted invoices.	Map extracted codes to ERPNext billable items dynamically.
Security	Consent, audit trails, restricted retention, and no silent auto-filing.	High-risk workflows need explicit human approval.

3. AI Model Strategy

Choose model roles instead of chasing one model for everything.

Layer	Recommended Role	Best Fit	Notes
Speech-to-text	Audio transcription	Whisper-class model or managed STT	Use domain noise filtering, speaker separation if possible.
Clinical summarization	SOAP generation	Strong reasoning model or tuned medical model	Needs strict JSON schema and hallucination controls.
Medical coding extraction	ICD/CPT suggestions	Dedicated prompt or smaller coding specialist model	Never trust coding output without validation.
Billing assistant	Item lookup + invoice draft	Application logic first, LLM second	LLM can suggest codes; ERPNext should decide billable items.

Practical rule: use the LLM to extract and propose, but use Frappe/ERPNext logic to validate and transact.

Open-source path

Self-hosted STT

Open-weight note-generation model

Local or private-cloud deployment

Best for data sovereignty

Hybrid path

Local audio capture

Managed reasoning model for SOAP

On-prem or VPC middleware

Best for implementation speed

Enterprise path

Private inference + governance

Human QA queues

Role-based approvals

Best for hospital-scale rollout

4. Reference Architecture

A practical Frappe Health-native workflow for ambient care documentation.

1. Audio Capture

Doctor-patient or nurse-patient conversation recorded through browser, tablet, or kiosk mic.

2. Transcription

Speech engine converts audio into timestamped text, optionally with speaker separation.

3. Structuring

LLM converts transcript into strict SOAP JSON, medications, diagnosis, and next-step suggestions.

4. Frappe Draft

Create Patient Encounter draft with subjective, objective, assessment, plan, and child tables.

5. Coding & Billing

Validate ICD/CPT and map approved billable items into ERPNext Sales Invoice draft.

6. Human Review

Doctor or billing user edits, signs off, submits, and closes the workflow.

Suggested custom DocTypes / app modules

DocType	Purpose
AI Scribe Session	Stores encounter audio metadata, transcript status, model used, and review state.
AI Transcript Chunk	Optional segmented transcript storage for long sessions.
AI Coding Suggestion	Stores proposed ICD/CPT/E&M with confidence and reviewer action.
AI Billing Draft Log	Tracks mapping of extracted codes to ERPNext items.
AI Audit Event	Captures access, generation, review, and submission events.

Recommended UI design inside Frappe

Encounter form action: “Start Ambient Note”

Live recorder panel: waveform, consent state, elapsed time, pause/resume

Transcript split view: raw transcript on left, structured SOAP draft on right

Coding assistant drawer: primary diagnosis, secondary diagnoses, CPT / procedure candidates

Billing panel: matched items, missing items, pricing, approval notes

Build for editability, not for one-click automation. Clinicians trust systems they can correct quickly.

5. Step-by-Step Implementation Plan

Start simple, then layer safety, coding, and billing.

Phase 1

Encounter draft creation

Record audio, transcribe, and push SOAP draft into Patient Encounter.

Phase 2

Medication & diagnosis extraction

Populate child tables for drugs and diagnosis suggestions.

Phase 3

Coding validation

Match model output with your Medical Code and Item master.

Phase 4

Invoice drafting

Create Sales Invoice drafts after human review and pricing rules.

Detailed implementation sequence

Create a dedicated API user in Frappe for the scribe service.
Generate API Key and API Secret and restrict permissions to only required DocTypes.
Add a custom app such as techunison_ai_scribe to isolate this feature cleanly.
Create an “AI Scribe Session” DocType for session tracking and traceability.
Add a custom button on Patient Encounter to launch audio capture.
Transcribe audio using your chosen speech layer.
Send transcript to the note-generation prompt and require JSON-only output.
Validate schema server-side before POSTing to Frappe Health.
Populate Encounter draft, including diagnosis and medication child tables where applicable.
Run ICD/CPT mapping logic against Frappe master data.
Build a billing preview and create a draft Sales Invoice.
Add review states: Generated, Reviewed, Approved, Rejected.
Log every generation, edit, submit, and invoice action in an audit table.

6. Source Code and Prompts

Starter examples for Techunison implementation. Review and adapt field names to your exact Frappe version and customizations.

6.1 Frappe REST API pattern for Patient Encounter

6.2 Prompt for SOAP JSON extraction

6.3 Prompt for ICD-10 extraction

6.4 Encounter with diagnosis + medication child tables

6.5 Sales Invoice draft creation

6.6 Automatic item mapping from extracted codes

6.7 Frappe custom app server stub

7. Security, Governance, and Clinical Safety

This is the part many teams under-design. Do not skip it.

Consent and transparency

Record patient consent for audio capture.

Display clear recording status in UI.

Store consent event in audit trail.

Access and retention

Role-based access for transcript and notes.

Limit raw audio retention aggressively.

Separate audio, transcript, and final note permissions.

Clinical control

Mandatory human review before submission.

Flag low-confidence outputs.

Never auto-finalize diagnoses or invoices silently.

Security checklist

Control	Implementation Guidance
API auth	Use token-based auth with dedicated integration user and least privilege.
Transport security	Use HTTPS/TLS for all API traffic and media upload flows.
Encryption	Encrypt stored media, transcripts, and related secrets at rest.
Audit trail	Track transcript creation, note generation, code edits, invoice draft creation, and submit actions.
Prompt safety	Require strict JSON output and reject malformed payloads server-side.
Data minimization	Store only what is operationally needed; purge raw audio when policy allows.
Review workflow	Save all AI-generated records as drafts first.
Master data validation	Validate ICD/CPT and item codes against local Frappe masters before billing.

8. Implementation Portfolio

A portfolio of implementations across industries, showcasing scalable digital solutions and workflow-driven system design.
Built to enable integration, automation, and efficient operations at scale.

Frappe Assistant Core: AI Execution Layer for ERPNext

MCP-based platform connecting ERPNext with GPT, Claude, and enterprise agents through secure, standardized workflows.

Frappe Assistant Core: LLM Integration Layer for ERPNext

Connects ERPNext to MCP-compatible AI models through secure, standardized tools, enabling AI-driven workflows and automation.

9. Delivery Roadmap for Techunison

A realistic path from pilot to hospital-scale deployment.

Stage 1: Proof of Value

Single specialty pilot

Encounter draft only

No auto-billing

Doctor review mandatory

Stage 2: Structured Coding

Diagnosis suggestions

Medication extraction

Item validation against masters

Billing preview dashboard

Stage 3: Revenue Workflow

Draft Sales Invoice creation

Billing desk approval

Exception queues

Audit and KPI reporting

Recommended KPIs

KPI	Why it matters
Time saved per encounter	Measures documentation efficiency.
Draft acceptance rate	Shows note quality and clinician trust.
Coding correction rate	Shows whether the model is safe for coding support.
Invoice draft match rate	Shows billing usefulness without over-automation.
Audio-to-note turnaround time	Important for busy OPD flow.
Human override frequency	High values reveal prompt or model weakness.

10. Research Sources

Source list included in the HTML for stakeholder review.

Frappe Framework REST API documentation
Frappe Framework token-based authentication documentation
Frappe Health Patient Encounter documentation
Frappe Health Patient documentation
scribeHC GitHub repository
Provided research notes consolidated from the uploaded conversation text

This blueprint intentionally treats model/vendor choice as replaceable, while keeping Frappe Health and ERPNext workflow ownership central.

Techunison implementation stance

The strongest version of this product is not “an AI note generator.” It is a clinical workflow automation layer built around Frappe Health, where ambient documentation, coding support, and billing assistance are integrated into one governed workflow.

Talk to an Expert

Request a Demo