Ajay
All projects

Voice-to-notes AI prototype

VoiceMed AI

A synthetic-data prototype that turns speech into a structured draft and blocks export until a person confirms it. It does not diagnose or prove clinical accuracy.

Building· GitHub — open VoiceMed AI repository

Test-grounded behavior

Replay the verified system path

Deterministic replay of linked tests—not live telemetry.

Expected test outcome

The download control remains disabled until the reviewer explicitly confirms the synthetic draft.

Synthetic browser-test replay—not a clinical workflow, accuracy result, or medical-device claim.

  1. Open the documentation labThe browser test enters the public console and starts the deterministic synthetic example.input
  2. Render an unreviewed draftThe interface labels the generated content as unreviewed and exposes the review state explicitly.processing
  3. Keep export lockedBefore confirmation, the Download reviewed draft control must remain disabled.failure
  4. Confirm and unlockAfter the human-review checkbox is selected, the same browser test requires the download control to become enabled.verified

Animation is paused. The Previous and Next controls keep every step available.

Inspect the pinned browser test

Human review gate: step 1 of 4. Open the documentation lab. The browser test enters the public console and starts the deterministic synthetic example.

16 / 16

unit and contract tests passing

Vitest covers the public synthetic fixture, bundled reference-data contracts, strict clinical schemas, rejected malformed output, and the deterministic provider contract.

Pinned unit and contract tests
Commit
824b233
Reproduce
npm run test
Environment
GitHub Actions ubuntu-latest · Node.js 22 · deterministic synthetic fixtures
Snapshot
Limit
Contract and fixture coverage only; no clinical-accuracy, diagnosis, or real-provider reliability claim.

4 / 4

desktop and mobile browser checks passing

Playwright verifies public safety copy and that export remains disabled until human review is explicit.

Pinned review-gate browser test
Commit
824b233
Reproduce
npm run build && npm run test:e2e
Environment
GitHub Actions ubuntu-latest · Node.js 22 · Chromium desktop and Pixel 7 emulation
Snapshot
Limit
Synthetic no-key browser workflow; it does not measure clinical accuracy or external-provider performance.

Problem

The engineering question

Clinical documentation is time consuming, while unconstrained generation creates unacceptable ambiguity. The prototype explores multilingual capture, schema-bound extraction, and mandatory review.

Constraints

What the design must respect

  • The public demo uses synthetic examples and must not retain health information.
  • Every generated report requires visible human review before export.
  • The product must state clearly that it does not diagnose or replace clinical judgment.

Architecture decisions

Deliberate choices and why they exist

01

Deterministic demo mode

A no-key path returns fixed synthetic fixtures so reviewers can inspect the complete workflow without transmitting data.

02

Schema before prose

Model output is validated as structured entities before it is rendered as a clinical note.

03

Review is a state

Draft, reviewed, and exported states are explicit rather than implied by a successful model response.

Failure modes

Designed to fail visibly

  • Unsupported language or low-quality audio.
  • Missing, malformed, or contradictory structured fields.
  • Provider timeout or rate limit.
  • A user attempting to export an unreviewed draft.

Verification

How claims are checked

  • Schema and API contract tests.
  • Deterministic no-key end-to-end browser flow.
  • One public synthetic fixture and bundled reference-data contracts.
  • No persisted audio or personal health information in the public demo.

Limitations

What remains unproven

  • The project is a research prototype and is not a medical device.
  • Accuracy metrics remain unpublished until the evaluation set and method are checked in.

Inspect the work

Source and evidence

The repository is the source of truth for implementation status. The case study never upgrades a planned feature into a shipped claim.

View repository