Burrito’s Labs Indie Tech Outmost Studio
Professional Systems
Active development

ClinScribe Local

A local-first ophthalmology documentation assistant that turns consultation audio into evidence-linked, review-required draft notes under physician control.

Category
Professional Systems
Platform
iPhone · iOS 18+
Status
Active development

What if clinical documentation assistance never obscured the boundary between evidence and inference?

Local sessions, consent-gated recording, supported on-device transcription, conservative candidate extraction, transcript evidence links, deterministic drafts, guardrails, and explicit review keep assistance inspectable.

User value

See which transcript passage led to each candidate statement.

Move from consultation audio to a review-required ophthalmology draft while keeping evidence, uncertainty, and physician judgment visible.

01

Retain local control

Keep the implemented session, audio, transcript, candidates, note, and review artifacts in one deletable app-container folder.

02

Trace every candidate

Open the transcript text and timestamp that produced a candidate instead of accepting an unsupported summary.

03

Review before reuse

Keep drafts and review prompts explicitly non-final so workflow actions never become clinical validation.

ClinScribe Local iPhone home screen showing an empty local consultation list, selected documentation language, privacy reminder, and New Consultation action.View full size
Current local workspaceCurrent simulator capture with an empty consultation list, selected language, privacy reminder, and local-only workflow entry points.

Features

A staged path from recording to review-required draft.

Each stage keeps source evidence, generated wording, and physician review separate.

  1. 01

    Conservative capability check

    Report device, operating-system, speech, language, and Foundation Models availability without inventing support when the runtime cannot verify it.

  2. 02

    Consent-gated local recording

    Create a local session, select one of five documentation languages, confirm consent, and record consultation audio as a local M4A file.

  3. 03

    Local transcript boundary

    Use Apple speech services only when supported and persist timestamped transcript segments inside the consultation session.

  4. 04

    Evidence-linked candidates

    Detect provisional vocabulary mentions, laterality, negation, and high-risk terms, then connect each candidate to transcript text and time ranges.

  5. 05

    Deterministic draft and guardrails

    Render fixed ophthalmology sections from candidates and check draft claims against available evidence with inspectable rules.

  6. 06

    Review prompts and audit history

    Turn guardrail findings into local prompts with status transitions and history without treating any action as clinical approval.

How it works

Capture, trace, draft, review.

The current implementation keeps the entire consultation pipeline inside one local session lifecycle.

  1. 01

    Check the device

    Review local capability results and choose the documentation language before creating a consultation.

  2. 02

    Create the local session

    Add an optional title or patient reference and keep the new session inside the app container.

  3. 03

    Confirm consent and record

    Pass the explicit consent transition, request microphone access, and capture a local audio file.

  4. 04

    Transcribe when supported

    Create timestamped transcript segments through the conservative Apple speech boundary.

  5. 05

    Extract candidates with evidence

    Identify bounded clinical mentions and retain the exact transcript segment and time range behind each one.

  6. 06

    Render and review the draft

    Generate deterministic sections, inspect guardrail prompts and audit state, then retain or delete the local session.

Current interface

Capability, consent, recording, and privacy stay visible.

These simulator captures use empty or synthetic local sessions and large system text. They are development evidence, not clinical validation or final App Store media.

ClinScribe Local capability screen showing device, iOS, speech, Foundation Models, and language checks with simulator-dependent items marked unknown.View full size
Current simulator capability report with conservative unknown states.

Capability check

The app reports what it can verify and leaves simulator-dependent speech and model capabilities visibly unknown.

ClinScribe Local New Consultation screen showing documentation language, optional session fields, privacy guidance, and consent controls.View full size
Current local session setup with no real patient data.

New consultation

Language, optional identifiers, local-processing copy, and the consent transition appear before recording begins.

ClinScribe Local consultation screen showing a synthetic session ready to record, microphone permission status, and local-only guidance.View full size
Synthetic development session before audio capture.

Recording workflow

The consultation screen keeps recording readiness, microphone permission, and local-only handling visible.

ClinScribe Local settings screen showing app version, language, consent, transcript, microphone, capability, and privacy controls.View full size
Current settings and privacy surface in the simulator build.

Privacy and settings

Language, consent, transcript, microphone, capability, and privacy settings remain available from one screen.

Product concept

Evidence provenance is not clinical truth.

The system can show where wording came from and where attention is needed without certifying a diagnosis, plan, or final note.

01

Candidate before fact

Extraction produces review-required candidates whose source is inspectable, not verified clinical assertions.

02

Draft before record

Generated sections remain working material and never become a final clinical record automatically.

03

Review state before validation

Acknowledged, deferred, dismissed, rejected, resolved, and reopened describe workflow state, not medical correctness.

Architecture

A local, staged, deterministic documentation pipeline.

A session folder owns the audio, transcript, candidates, draft, prompts, and audit artifacts from creation through deletion.

  1. 01Local consultation session
  2. 02Consent-linked M4A recording
  3. 03Timestamped transcript segments
  4. 04Evidence-linked candidate facts
  5. 05Deterministic section draft
  6. 06Guardrails, prompts, and deletion

The session owns its lifecycle

Application Support folders and local JSON keep related artifacts together and give deletion one bounded target.

Rules remain inspectable

Candidate extraction, section rendering, claim checks, and prompt mapping use explicit deterministic logic.

Model generation stays disabled

Foundation Models are represented as an availability and future integration boundary, not a production note generator.

Privacy claims remain bounded

Managed paths are excluded from backup, while physical-device file protection, uninstall behavior, and deletion completeness still require verification.

Technology

Built as an inspectable local workflow.

Application

  • Swift 6
  • SwiftUI
  • iPhone
  • iOS 18+

Capture and speech

  • AVFoundation
  • Local M4A
  • Apple Speech APIs
  • Availability checks

Clinical pipeline

  • Deterministic candidates
  • Laterality and negation
  • Evidence links
  • Draft guardrails

Local data

  • Application Support folders
  • Atomic JSON
  • Backup exclusion
  • Session deletion

Current boundary

  • XCTest
  • No server or account
  • No analytics SDK
  • No EHR transfer
  • Model note generation disabled

Development

The review pipeline is implemented. Clinical and device evidence remain pending.

The SwiftUI workflow, local persistence, recording boundary, transcript models, deterministic extraction, evidence links, draft rendering, guardrails, review prompts, audit actions, and deletion are implemented and test-backed. Real-device speech, clinical usefulness, multilingual safety copy, storage protection, accessibility, legal, regulatory, and release readiness remain unvalidated.

Stage
Active development
Unit tests
533 passed
UI tests
7 passed
Clinical validation
Not completed

Implemented

  • SwiftUI consultation list, session creation, consent, recording, review, settings, and local deletion surfaces
  • Application Support session folders, atomic JSON artifacts, local M4A capture, and backup exclusion
  • Conservative Apple speech and language capability boundaries
  • Deterministic vocabulary, laterality, negation, and high-risk candidate extraction with transcript evidence links
  • Deterministic section rendering, claim guardrails, review prompts, status transitions, and audit history
  • Automated unit and UI coverage using empty, synthetic, and injected test states

Before clinical or public use

Complete manual and physical-device QA, obtain ophthalmologist and native-speaker review, verify storage and deletion behavior, resolve terminology governance and stale copy, complete legal, privacy, medical-regulatory, accessibility, and App Store review, and separately design any copy, export, EHR, or model-assisted generation path.

ClinScribe Local lab notes

The project, over time.

Notes about the thinking, architecture, development, and decisions behind this work.

  1. ClinScribe Local
    21 Jun 2026 · Draft

    Turning Guardrails Into a Review Workflow

    Draft issues became local prompts with status and audit history without becoming clinical validation.
  2. ClinScribe Local
    18 Jun 2026 · Draft

    From Transcript Mentions to Evidence-Linked Drafts

    Deterministic extraction, source links, guarded wording, and claim checks made traceability the central product mechanism.
  3. ClinScribe Local
    16 Jun 2026 · Draft

    A Local-First Foundation

    The first implementation made each consultation a local, deletable session and every generated artifact review-required.