OpenMRS LiveKit Voice Assistant

OpenMRS LiveKit is a local-first clinical voice assistant for OpenMRS O3. It opens a patient-scoped LiveKit audio room from the chart, supports a doctor-patient voice workflow, redacts PHI-like text, and produces a structured OpenMRS encounter draft for clinician review.

The project is designed for clinics where internet connectivity is unreliable, privacy matters, and bilingual encounters are common.

At a Glance

| Area | Current behavior | | --------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Frontend | OpenMRS O3 microfrontend with Carbon UI, LiveKit room controls, patient context, privacy status, transcript, and draft review. | | Helper service | Local Python service for LiveKit tokens, health/readiness, PHI redaction, synthetic consultations, encounter compilation, draft queueing, and optional OpenMRS writes. | | Realtime agent | Companion repository: sihsalus/openmrs-livekit. It owns STT, LLM/tool calls, TTS, data-channel events, and OpenMRS draft events. | | Safety boundary | Assistive documentation only. Drafts are queued for clinician review and are not written to OpenMRS unless explicitly configured and requested. | | Synthetic data | Deterministic synthetic consultations for validation and end-to-end checks without real patient data. |

What It Does

• Starts a patient-scoped LiveKit audio room from an OpenMRS O3 extension.
• Derives safe defaults for LiveKit URL, token endpoint, room prefix, and clinical language metadata.
• Supports local-first speech-to-text, clinical translation, text-to-speech, and encounter drafting through the LiveKit agent and helper service.
• Consumes agent-data LiveKit data-channel messages for agent status, transcript payloads, clinical facts, and drafts.
• Redacts PHI-like identifiers before generated text is displayed or queued.
• Builds reviewable drafts with chief complaint, symptoms, medications, allergies, assessment notes, patient instructions, facts, missing fields, and review queue items.
• Avoids storing raw audio by default.
• Queues drafts locally and can optionally write OpenMRS encounter payloads only when write mode, credentials, and OpenMRS metadata are configured.

Architecture

flowchart LR
  chart[OpenMRS O3 patient chart] --> mfe[esm-livekit-app]
  mfe --> room[LiveKit room]
  room --> agent[LiveKit AI agent]
  agent --> helper[local helper service]
  helper --> draft[redacted encounter draft]
  draft --> review[clinician review queue]
  review --> openmrs[(optional OpenMRS encounter write)]

  helper -. health/token/draft .-> mfe
  agent -. agent-data topic .-> mfe

OpenMRS O3 patient chart
  -> LiveKit room
  -> LiveKit AI agent
  -> local helper service
  -> configured model provider or Ollama-compatible drafting
  -> redacted encounter draft
  -> clinician review queue

Repository layout:

| Path | Purpose | | ------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------- | | src/ | OpenMRS O3 microfrontend, LiveKit token client, room UI, patient context, agent data parsing, draft review UI, and tests. | | token-server/ | Local helper service, local AI contracts, synthetic data, redaction, draft queue/write bridge, smoke tests, and e2e contract tests. | | deploy/openmrs-base-livekit/ | Reproducible OpenMRS base stack integration with LiveKit server, CPU agent, helper service, gateway routes, CSP, and frontend config. | | translations/ | English and Spanish UI messages. |

Frontend source is grouped by domain:

| Path | Purpose | | --------------- | ---------------------------------------------------------------------------------------- | | src/voice/ | OpenMRS voice action, modal, consultation panel, and voice panel styles. | | src/audio/ | Audio visualizer and microphone availability/control helpers. | | src/livekit/ | LiveKit token client, agent data-channel parser, helper health normalization, and tests. | | src/patient/ | Patient context loading, demographics normalization, patient context UI, and tests. | | src/clinical/ | Clinical language defaults and encounter draft merging/materialization. |

AI Model Boundary

The browser frontend does not hardcode or run a foundation model. It connects the OpenMRS chart to LiveKit, sends room metadata, and consumes agent data-channel events. Model provider selection belongs to the companion agent and helper configuration.

Known defaults in this repository:

• Standalone helper /compile-encounter: OLLAMA_MODEL defaults to medgemma:latest and falls back to deterministic heuristics when Ollama is not reachable.
• OpenMRS base deployment: the helper and CPU agent default OLLAMA_MODEL to qwen2.5:1.5b unless overridden in the deployment environment.
• OpenMRS base deployment sets the CPU agent LLM_PROVIDER=ollama.

The base agent prompt lives in the companion agent repository. The current reference configuration targets Spanish clinical encounters in a Latin American OpenMRS setting and can be replaced by site-specific session instructions in the agent layer.

Clinical Safety and Privacy

| Control | Implementation posture | | ---------------- | --------------------------------------------------------------------------------------------------------------- | | Raw audio | Not stored by default. Recording requires explicit consent workflow and storage controls. | | PHI-like text | Deterministic redaction for helper-generated text and synthetic identifiers. | | Drafts | Clinician-reviewable queue by default, not autonomous charting. | | OpenMRS writes | Disabled by default. Requires explicit write request plus enabled server configuration. | | Audit events | Draft lifecycle events exclude transcript and draft text; patient references are hashed. | | Local-first mode | Supported for offline-capable deployments through local LiveKit, agent, helper, and Ollama-compatible drafting. |

This project is not a diagnostic system and does not make autonomous clinical decisions. Generated output must be reviewed by a clinician before charting.

Quick Start

Install frontend dependencies:

yarn install

Run the OpenMRS frontend:

yarn start

Install and run the local helper:

python3 -m venv token-server/.venv
token-server/.venv/bin/pip install -r token-server/requirements.txt
LIVEKIT_API_KEY=<key> LIVEKIT_API_SECRET=<secret> token-server/.venv/bin/python token-server/server.py

The standalone helper listens on port 7890 by default. The OpenMRS base deployment does not expose that port on the host; browser traffic should go through the gateway. Base deployments use /openmrs/livekit/token so OpenMRS session cookies scoped to /openmrs can reach the helper when session validation is enabled. When OpenMRS config is left blank, the frontend derives standalone development endpoints:

| Setting | Derived default | | ----------------- | --------------------------------------------- | | LiveKit WebSocket | ws(s)://<current-browser-host>:7880 | | Token endpoint | http(s)://<current-browser-host>:7890/token | | Room prefix | openmrs-voice- |

For any shared evaluation, staging, or production deployment, enable the readiness gate and configure browser origins explicitly:

TOKEN_SERVER_ENV=production
TOKEN_SERVER_ALLOWED_ORIGINS=https://openmrs.example.org
LIVEKIT_API_KEY=<site-livekit-api-key>
LIVEKIT_API_SECRET=<site-livekit-api-secret>

Production mode fails fast if LiveKit signing credentials or the CORS allowlist are missing. The helper also creates local draft, recording manifest, and audit JSONL files with owner-only permissions (0600). That is useful for a controlled evaluation host, but it is not a substitute for encrypted storage in a regulated deployment.

Configuration

The OpenMRS module config schema exposes:

| Config key | Meaning | | ------------------ | ---------------------------------------------------------------------------------------------------- | | livekitServerUrl | LiveKit WebSocket URL. Leave blank to derive the browser-host default. | | tokenEndpoint | Helper endpoint used to request LiveKit room tokens. Leave blank to derive the browser-host default. | | roomPrefix | LiveKit room prefix joined by the local agent. Defaults to openmrs-voice-. |

At consultation start, the microfrontend derives default clinician, patient, and agent voice languages from the active OpenMRS locale. English is the default unless the OpenMRS locale is Spanish (es, es-PE, es_MX, etc.). The user can adjust the values before the room is created.

The token request sends:

{
  "patientUuid": "aefc6e8d-fdc7-430f-9dae-a1dcbff2cdec",
  "roomName": "openmrs-voice-aefc6e8d-fdc7-430f-9dae-a1dcbff2cdec",
  "roomPrefix": "openmrs-voice-",
  "doctorLanguage": "es",
  "patientLanguage": "en",
  "agentVoiceLanguage": "es",
  "captureRole": "doctor",
  "defaultHumanRole": "doctor",
  "speakerAttributionMode": "source-role"
}

The helper normalizes doctorLanguage, patientLanguage, and agentVoiceLanguage to supported base codes (en and es). Unsupported patient languages fall back to the normalized clinician language. If LIVEKIT_HTTP_URL is configured, the helper performs a best-effort LiveKit room metadata sync before returning the browser token.

The browser microphone is requested as captureRole=doctor. If the STT provider emits a speaker_id, the agent can include speakerId, attribution mode, and attribution source in transcript payloads. If no speaker ID is available, the transcript falls back to the configured default human role instead of claiming automatic doctor/patient diarization.

Helper Service Contracts

The helper is not the realtime conversational agent. It provides local contracts that support the frontend, validation workflows, and smoke tests.

| Endpoint | Contract | | ------------------------------ | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | GET /health | Reports LiveKit, OpenMRS, Ollama, agent, STT/TTS, token signing, CORS, local storage, draft audit, and production readiness status. | | POST /token | Returns an HS256 LiveKit JWT with a room-scoped join grant and optional room metadata sync. | | POST /compile-encounter | Requires transcript or text, redacts PHI-like text, and compiles a clinician-reviewable OpenMRS draft using Ollama when available or deterministic heuristics otherwise. | | POST /synthetic-consultation | Generates deterministic synthetic dialogue, redacted transcript, draft, and an openmrsDraftRequest for validation and e2e tests. | | POST /recording/session | Records a consent manifest for future recording workflow; it does not capture or store raw audio by default. | | POST /openmrs/draft | Queues a draft locally by default and can optionally create an OpenMRS encounter through /openmrs/ws/rest/v1. |

Important helper environment variables:

| Variable | Use | | ---------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------- | | LIVEKIT_API_KEY, LIVEKIT_API_SECRET | LiveKit token signing. Missing values fall back to LiveKit dev defaults only in development mode. | | LIVEKIT_HTTP_URL | Optional LiveKit HTTP API URL for best-effort room metadata sync. | | LIVEKIT_ROOM_PREFIX | Room prefix shared by frontend, helper, and agent. | | TOKEN_SERVER_ENV | Set to production for shared evaluation and production-like deployments. | | TOKEN_SERVER_REQUIRE_PRODUCTION_CONFIG | Enables production readiness checks without changing TOKEN_SERVER_ENV. | | TOKEN_SERVER_REQUIRE_OPENMRS_SESSION | Requires protected helper endpoints to validate the caller's OpenMRS session cookie or Authorization header. Required by production readiness. | | TOKEN_SERVER_ALLOWED_ORIGINS | Comma-separated browser origins accepted by CORS. Required for production readiness. | | OLLAMA_MODEL | Model name used by helper /compile-encounter when Ollama is available. | | DRAFT_STORE_PATH | Local JSONL draft queue path. | | RECORDING_MANIFEST_PATH | Local JSONL recording consent manifest path. | | AUDIT_LOG_PATH | Draft lifecycle audit JSONL path. | | AUDIT_HASH_SALT | Site-managed salt for hashed patient references in audit events. |

Optional OpenMRS write configuration:

OPENMRS_DRAFT_WRITE_ENABLED=true
OPENMRS_ENCOUNTER_TYPE_UUID=<encounter-type-uuid>
OPENMRS_LOCATION_UUID=<location-uuid>
OPENMRS_DRAFT_OBS_CONCEPT_UUID=<text-concept-uuid-for-ai-draft>
OPENMRS_PROVIDER_UUID=<provider-uuid>
OPENMRS_ENCOUNTER_ROLE_UUID=<encounter-role-uuid>
OPENMRS_STRUCTURED_OBS_CONCEPTS='{"chiefComplaint":"...","symptoms":"...","medicationsMentioned":"...","allergiesMentioned":"...","assessmentNotes":"...","patientInstructions":"..."}'

To request a real OpenMRS write, send writeToOpenmrs: true or mode: "write". The server writes only when OPENMRS_DRAFT_WRITE_ENABLED=true and required OpenMRS metadata are configured. Real writes also require a visitUuid for an active visit belonging to the patient; drafts without an active visit remain queued for review. Authentication can come from OPENMRS_USERNAME and OPENMRS_PASSWORD, OPENMRS_BASIC_AUTH, or the forwarded OpenMRS browser session cookie.

Draft lifecycle audit event types:

• draft_queued: draft was queued locally without an OpenMRS write.
• draft_saved: OpenMRS accepted the encounter create request.
• draft_write_rejected: an OpenMRS write was requested but blocked or rejected by configuration, authentication, patient lookup, or the OpenMRS REST API.

Audit events intentionally do not store raw transcript or draft text. The local draft queue stores the redacted transcript and clinician-reviewable draft so the UI can recover queued work; treat DRAFT_STORE_PATH as clinical data even after redaction.

OpenMRS Base Deployment

The reproducible OpenMRS base stack lives in deploy/openmrs-base-livekit. It adds the LiveKit server, CPU agent, helper/token service, gateway routes, CSP, and frontend module configuration without committing site secrets.

Set these values in the OpenMRS distro .env or export them before running Docker Compose:

OPENMRS_DISTRO_ROOT=/path/to/openmrs-distro-referenceapplication
OPENMRS_ESM_LIVEKIT_PATH=/path/to/openmrs-esm-livekit
OPENMRS_LIVEKIT_AGENT_PATH=/path/to/openmrs-livekit
LIVEKIT_HOST=<browser-reachable-host>
OPENMRS_LIVEKIT_SERVER_URL=<optional-browser-wss-livekit-url>
LIVEKIT_API_KEY=<site-livekit-api-key>
LIVEKIT_API_SECRET=<site-livekit-api-secret>
AUDIT_HASH_SALT=<site-managed-random-salt>
OPENMRS_PASSWORD=<openmrs-admin-password>

The normal frontend deployment path is npm:

npm view @sihsalus/esm-livekit-app version

Set OPENMRS_LIVEKIT_FRONTEND_VERSION only to a version that is actually published on npm. If a release tag builds successfully but npm publish fails, the OpenMRS frontend can temporarily serve a locally built dist/ directory via the importmap and route registry for deployment recovery, but that hotfix is not the long-term reproducible path.

Install into the OpenMRS distro:

mkdir -p "$OPENMRS_DISTRO_ROOT/deploy/livekit"
cp deploy/openmrs-base-livekit/*.yml "$OPENMRS_DISTRO_ROOT/deploy/livekit/"
cp deploy/openmrs-base-livekit/*.Dockerfile "$OPENMRS_DISTRO_ROOT/deploy/livekit/"

OPENMRS_DISTRO_ROOT="$OPENMRS_DISTRO_ROOT" \
LIVEKIT_HOST="$LIVEKIT_HOST" \
python3 deploy/openmrs-base-livekit/configure_base_livekit.py

Run from the OpenMRS distro root:

docker compose \
  -f docker-compose.yml \
  -f deploy/livekit/build.yml \
  -f deploy/livekit/livekit.yml \
  up -d --build

Verify:

docker compose ps
curl http://<openmrs-host>/openmrs/livekit/health
docker logs openmrs-distro-referenceapplication-livekit-helper-1
docker logs openmrs-distro-referenceapplication-livekit-agent-cpu-1

For HTTPS deployments, route LiveKit through the gateway WebSocket proxy and use a browser-trusted hostname:

SSL_MODE=dev
CERT_WEB_DOMAINS=openmrs.example.org,localhost
CERT_WEB_DOMAIN_COMMON_NAME=openmrs.example.org
OPENMRS_LIVEKIT_SERVER_URL=wss://openmrs.example.org/livekit-sfu
TOKEN_SERVER_ALLOWED_ORIGINS=https://openmrs.example.org,http://openmrs.example.org

The deployed bundle should include the OpenMRS base FHIR workaround: it fetches MedicationRequest?patient=<uuid>&_count=20 and filters active medication requests locally instead of sending status=active to the base distro FHIR endpoint.

Tests

Run the frontend test suite:

yarn test

Run the helper contract tests:

yarn test:e2e:token-server

The helper e2e test starts fake local OpenMRS, Ollama, and LiveKit services, then validates health checks, PHI redaction, synthetic data generation, recording consent, CORS, and an authenticated OpenMRS encounter write against the fake REST API.

Run a smoke test against a real running helper:

yarn test:smoke:token-server

For a remote helper:

TOKEN_SERVER_SMOKE_URL=https://helper.example.org yarn test:smoke:token-server

End-to-End Smoke Test

This checklist is the minimum bar before presenting the prototype as tested in a real environment. It complements automated unit and contract tests; it does not make the project production-ready by itself.

Automated Helper Smoke

Start the helper with the same environment used by the target deployment:

LIVEKIT_API_KEY=<key> LIVEKIT_API_SECRET=<secret> python3 token-server/server.py

Run:

yarn test:smoke:token-server

The smoke test verifies:

• /health reports the agent-data data-channel contract.
• /token returns an HS256 LiveKit JWT with a room-scoped join grant.
• /compile-encounter redacts name, email, phone, local document IDs, and OpenMRS ID values.
• /synthetic-consultation returns synthetic, redacted consultation data.
• /openmrs/draft queues a clinician-reviewed draft without writing to OpenMRS.

Real Environment Preflight

Use only synthetic patient data. Record these values before the browser smoke:

OpenMRS URL:
LiveKit WebSocket URL:
Token endpoint:
Agent command/container:
Room prefix:
Synthetic patient UUID:
OpenMRS encounter type UUID:
OpenMRS location UUID:
Draft obs concept UUID:

Required preflight checks:

1. OpenMRS is served over HTTPS, or every service is localhost-only.
2. livekitServerUrl is wss:// for shared environments.
3. tokenEndpoint is same-origin /openmrs/livekit/token or https:// for shared environments.
4. Helper runs with TOKEN_SERVER_ENV=production or TOKEN_SERVER_REQUIRE_PRODUCTION_CONFIG=true for shared evaluations.
5. Helper /health shows configured LiveKit token signing, enforced OpenMRS session validation, and a non-permissive CORS allowlist for the OpenMRS browser origin.
6. Frontend roomPrefix, helper LIVEKIT_ROOM_PREFIX, and agent LIVEKIT_ROOM_PREFIX match exactly.
7. Agent process is running with the intended STT, LLM, and TTS providers.
8. Browser microphone permission is granted and visible in the browser site settings.

Manual Browser Smoke

1. Open a synthetic patient chart and launch the LiveKit voice panel.

2. Confirm the health panel shows LiveKit, token server, local storage, agent, OpenMRS, and draft write readiness.

3. Confirm the agent publishes an agent_connected or agent_listening status on the agent-data data-channel topic before the first transcript.

4. Speak this synthetic utterance through the browser microphone:

   Paciente: Maria Fernanda Quispe, H.C. A-998877, vive en Av. Los Incas 123.
   Tiene tos seca desde hace cinco dias. Niega alergias a medicamentos.
   Toma paracetamol 500 mg cada ocho horas.

5. Confirm the live transcript arrives on the agent-data data-channel topic.

6. Confirm patient identifiers are redacted before display or draft persistence: Maria Fernanda Quispe, A-998877, and Av. Los Incas 123 must not appear in frontend transcript text, stored evidence snippets, queued draft text, or helper logs.

7. Confirm negation is preserved: niega alergias must not become a positive allergy.

8. Confirm medication and dose are preserved for review: paracetamol 500 mg cada ocho horas.

9. Confirm the draft shows missing fields and review queue items.

10. Queue the draft and verify no OpenMRS write occurs unless explicitly enabled.

11. Confirm the helper writes a draft_queued, draft_saved, or draft_write_rejected audit event without transcript or draft text.

12. If write mode is enabled, verify the created encounter uses the expected patient, encounter type, location, provider, role, and concept UUIDs.

13. Reload the patient chart and verify the saved or queued draft state is explainable to a clinician reviewer.

Deployment Logs

For the self-hosted stack, use container logs to verify the browser, LiveKit, helper, and agent path without storing raw clinical audio or transcript text:

docker logs -f openmrs-distro-referenceapplication-gateway-1
docker logs -f openmrs-distro-referenceapplication-livekit-helper-1
docker logs -f openmrs-distro-referenceapplication-livekit-1
docker logs -f openmrs-distro-referenceapplication-livekit-agent-cpu-1
docker logs -f openmrs-distro-referenceapplication-backend-1

Useful signals:

• Gateway: static microfrontend chunks, /openmrs/livekit/token, /openmrs/livekit/health, compatibility /livekit/* routes, and OpenMRS REST/FHIR status codes.
• Helper: token, health, synthetic consultation, compile, and draft queue requests.
• LiveKit: browser participant joins, agent assignment, ICE/UDP connection type, track publication, and room close reason.
• Agent: room connection, metadata parsing, prompt budgeting, readiness status, TTS/STT/LLM timing, and transcript-save policy.

The expected deployment logging posture is metadata and operational status only. Helper and agent logs must not include raw transcript text, draft text, or unredacted patient identifiers.

Known Weakness Validation

Room metadata validation:

docker logs openmrs-distro-referenceapplication-livekit-helper-1 \
  | grep -E "LiveKit room metadata (created|updated)"
docker logs openmrs-distro-referenceapplication-livekit-agent-cpu-1 \
  | grep -E "Metadata parsed|Room metadata derived from room name|Room metadata empty|Sending initial greeting"

Expected result:

• Helper logs LiveKit room metadata created or updated for rooms opened from the OpenMRS microfrontend when LIVEKIT_HTTP_URL is configured.
• Agent logs Metadata parsed when LiveKit room metadata is available.
• The helper room metadata includes normalized doctorLanguage, patientLanguage, agentVoiceLanguage, languageMode, speakerAttributionMode, and defaultHumanRole.
• English is the expected default when OpenMRS does not expose a Spanish locale. Spanish OpenMRS locales such as es, es-PE, or es_MX should produce Spanish room metadata.
• The agent uses doctorLanguage for STT language hints, agentVoiceLanguage for the initial greeting and assistant transcript language labels, and patientLanguage for patient-facing translation context.
• Agent may log Room metadata derived from room name as a non-blocking fallback for rooms named with the configured prefix, for example openmrs-voice-<patientUuid>.
• Room metadata empty should only appear for rooms that do not match the configured agent room prefix or cannot expose a patient UUID safely.
• Transcript payloads should include speakerId and attributionMode=stt-speaker-id when the STT provider emits speaker IDs. If no speaker ID is available, the payload should include attributionSource=missing-speaker-id and fall back to defaultHumanRole. That fallback is intentionally not presented as automatic diarization.

OpenMRS base FHIR MedicationRequest validation:

curl -I "$OPENMRS_BASE_URL/ws/fhir2/R4/MedicationRequest?patient=<uuid>&_count=20"
curl -I "$OPENMRS_BASE_URL/ws/fhir2/R4/MedicationRequest?patient=<uuid>&status=active&_count=20"

Expected result:

• The first request should return 200.
• On the observed OpenMRS base distro with fhir2-api-4.1.0, the second request can return 500 due to a backend NullPointerException.
• The microfrontend avoids that backend bug by fetching MedicationRequest without the status search parameter and filtering status === "active" locally.

Go / No-Go Criteria

Ready for supervised evaluation only if all are true:

• Browser joins the room without mixed-content errors.
• Microphone publishes audio and the agent receives it.
• Agent publishes readiness status over agent-data.
• At least one transcript and one draft arrive in the frontend.
• Synthetic identifiers are redacted before display or persistence.
• Draft remains reviewable and does not write to OpenMRS unless explicitly enabled.
• Draft lifecycle audit events are present and exclude raw transcript/draft content.
• If OpenMRS write is enabled, the encounter appears under the synthetic patient with the configured metadata and concept UUIDs.

No-go if any are true:

• Browser requires cleartext ws:// or http:// outside localhost.
• Token server accepts an unexpected browser origin in shared environments.
• Agent joins a different room prefix than the frontend.
• Raw synthetic identifiers appear in transcript, draft, logs, or JSONL queues.
• Audit JSONL events contain transcript text or draft text.
• niega alergias becomes a positive allergy.
• OpenMRS write occurs without explicit operator action and configuration.

Not Covered

• Browser-to-agent media quality across real clinic networks.
• LiveKit SFU TLS termination and certificate rotation.
• Application-level end-to-end media encryption.
• Encryption at rest for queued drafts, transcripts, logs, or recording manifests.
• Full OpenMRS role-based access review.
• Clinical validation by a clinician.
• Local clinical NER with site dictionaries. Current redaction is deterministic pattern matching with local Spanish healthcare identifiers.

Competitive Review

This review covers adjacent open-source medical scribe projects. It is a technical positioning exercise, not a claim that any project has been clinically validated or is production-ready for regulated care without site-specific security, privacy, and clinical governance review.

OpenMRS LiveKit is not trying to become a generic commercial ambient scribe. Its specific product wedge is:

OpenMRS O3 patient chart
  -> LiveKit realtime room
  -> local-first AI workflow
  -> PHI-like redaction
  -> evidence-backed encounter draft
  -> clinician review before OpenMRS write

Snapshot reviewed on 2026-07-06:

| Project | License | Competitive assessment | Relevance to OpenMRS LiveKit | | ------------------------------------------------------------------------- | ---------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------- | | Berta AI Scribe | Apache 2.0 | Strong deployment benchmark: FastAPI + Next.js, auth, local/cloud models, storage options, AWS path, and security posture. It is not OpenMRS-first. | Reuse deployment, auth, storage, and operations patterns. Avoid copying product scope wholesale. | | Open Medical Scribe | MIT | Strong modular product benchmark: swappable STT/LLM providers, local/cloud/hybrid operation, streaming, FHIR DocumentReference export, and multiple note types. | Reuse provider interfaces, audit events, local/cloud switching, note-quality workflows, and future FHIR export ideas. | | scribeHC | MIT | Useful workflow benchmark: Expo mobile capture, Next.js dashboard, FastAPI processing, authentication, and SOAP note editing. It is a parallel app workflow. | Use as a UX reference for recording, queue review, and editing. Keep this project embedded in the OpenMRS O3 patient chart. | | AI-Scribe | GPL-3.0 | Useful local-first reference: Python client/server, Whisper/Kobold-style processing, PHI scrubbing additions, and single-machine setup. Less modular. | Reference local processing patterns and failure modes only. Do not copy GPL-3.0 code into this MPL-2.0 repository. |

Positioning conclusion:

• Berta AI Scribe and Open Medical Scribe are credible technical benchmarks.
• Open Medical Scribe is the strongest source of reusable product architecture ideas, especially provider boundaries, note formats, and FHIR export.
• Berta AI Scribe is the stronger deployment and operations benchmark.
• scribeHC is useful for mobile capture and dashboard workflow patterns, but is not as close to the OpenMRS O3 embedded use case.
• AI-Scribe validates demand for local-first scribing, but is not a direct architecture target for this codebase.

Adopted now:

• Keep the current OpenMRS O3 frontend and LiveKit agent architecture.
• Keep local-first provider configuration instead of depending on one hosted model.
• Add helper-side draft lifecycle audit events for draft_queued, draft_saved, and draft_write_rejected.
• Keep audit events minimal: no transcript text, no draft text, and only hashed patient references.
• Keep the clinical safety claim narrow: assistive documentation with clinician review, not diagnosis or autonomous charting.

Later product work:

• Add a formal note-quality evaluation rubric, such as PDQI-9-style review, for clinician scoring of generated drafts.
• Add specialty-specific templates and OpenMRS concept mapping packs.
• Add local clinical NER with site dictionaries and a labeled Spanish PHI test corpus.
• Add encrypted storage and immutable audit storage for regulated deployments.
• Add optional FHIR export where OpenMRS implementations prefer FHIR resources over REST encounter payloads.

Clinical Use Case

The primary use case is point-of-care voice support inside the OpenMRS O3 patient chart. The system can run with local AI services, generate synthetic bilingual consultations for validation, redact patient identifiers, and produce a reviewable OpenMRS encounter draft.

The product focus is offline-capable clinical documentation assistance: clinician-controlled audio capture, translation support, PHI-aware transcript handling, evidence-backed drafts, and explicit review before any OpenMRS write.