01
AI that runs on the device
Speech recognition and note drafting run on the Neural Engine and GPU of your iPhone or Mac. There is no inference server, no third-party AI vendor and no cloud fallback. If a model can’t run, Care says so.
Scribed Care In development
The AI scribe that
never leaves the room.
Scribed Care records, transcribes and drafts encounter notes with AI that runs entirely on your iPhone or Mac. No cloud inference. No audio upload. Patient data stays in an encrypted vault that only you can open.
iPhone & Apple-silicon Mac · Online setup, then offline work
SCRIBED CareIllustrative preview · Synthetic data
Example person
A moment to listen.
Test encounter · On this device
Example recording Recording complete
Original transcript
“Let’s review them together.”
Ready to review
The conversation stays close.
From conversation to context
Every note has a source.
Keep the words and the draft in view.
Test encounterSource transcript
Visitor
I’d like to go over my notes from our last visit.
Clinician
Let’s review them together.
Original wording remains available while notes are reviewed.
Test encounterAI draft · Needs review
Visit context
The visitor asked to review notes from the previous visit. The clinician offered to review them together.
Based on the two example transcript passages
Illustrative documentation only. The draft requires human review against its source.
Explore the transcript and note draft above.
AI on the device, not a server Audio never leaves the device Encrypted vault · Face ID Works in airplane mode
What Care does
Ambient documentation.
Without the data trail.
Everything a medical scribe does, from the recording to the reviewed note, on hardware already in your pocket.
02
An encrypted vault, not a cloud folder
Patients, recordings, transcripts and notes are stored in an encrypted database whose key is held in the device’s protected keychain and released only by Face ID, Touch ID or your passcode.
03
Built for the exam room
Confirm the patient and recording consent, then record. Care transcribes on the device, flags medication names against RxNorm and keeps the original transcript beside every edit.
04
A draft. Then your judgment.
The local assistant drafts a note from the transcript. You verify each medication, check the source and finalize. Nothing is signed off without you.
On-premises, literally
Your exam room
is the data center.
Most AI scribes stream the visit to a vendor’s cloud and a third-party model. Care runs speech recognition and note drafting on the device in your hand, so patient audio has nowhere to go.
- Exam-room audio
- Uploaded over the internet
- Vendor servers
- Third-party AI model
Patient audio crosses the internet and sits on infrastructure you don’t control.
Inside your iPhone or Mac
- MicrophoneCaptured on the device
- Neural EngineTranscribed and drafted on the device
- Encrypted vaultPatients, recordings, transcripts, notes
- Patients & encounters
- Recordings & transcripts
- Notes & review history
No network needed after setup. No inference vendor in the path of PHI.
- 0Cloud AI calls during a visit
- 100%Of transcription and drafting on the device
- 1Owner per vault, unlocked by Face ID or Touch ID
Privacy and security
Defense in depth,
in the palm of your hand.
Keeping data on the device is the first layer. Here’s everything else between a patient’s words and anyone who shouldn’t hear them.
Encrypted at rest
Every patient, recording, transcript and note lives in an encrypted SQLCipher database. Temporary audio files are purged when the vault locks.
Keys that never leave the device
The vault key sits in the device’s protected keychain, bound to its passcode and this device only. It is never synced, backed up or sent anywhere.
Face ID, Touch ID, auto-lock
Opening the vault takes device authentication. Leave the app and it locks; the screen is covered in the app switcher.
No cloud fallback. Ever.
If a model is missing or a device runs out of memory, Care stops and tells you. It never quietly sends audio to a server instead.
No leaky exits
No share sheet, no clipboard, no analytics, no crash reporting in the clinical path. Exports go only where you save them.
Out of consumer backups
Clinical data and keys are excluded from iCloud and device backups. A passphrase-protected AES-256 vault archive is your recovery copy.
An append-only record
The original transcript is never overwritten. Every correction is a separate layer, hashed and traceable back to the source.
Verified models
Each model file is checked against a pinned checksum before it can load, so the AI that runs is the AI we tested.
HIPAA
Built for your HIPAA program.
Not a badge on a website.
HIPAA compliance belongs to your practice: your risk analysis, policies, training and agreements. What an app can do is shrink the surface you have to protect. When transcription and AI never leave the device, there is no cloud processor holding your patients’ audio.
- No AI vendor or transcription service receives PHI
- Cloud features off by default, opt-in per workspace
- Exports and archives go only where you save them
Scribed does not yet offer a business associate agreement, so the optional patient-list sync is unavailable to covered entities. Care’s on-device workflow doesn’t need it.
Security Rule technical safeguards, supported on the device
- Access control§164.312(a)
- One vault, one owner. Device authentication on every unlock and automatic locking.
- Audit controls§164.312(b)
- A local audit log of vault events, plus an append-only raw record behind every note.
- Integrity§164.312(c)
- Hashed raw transcripts and audio; edits are layers on top, never replacements.
- Transmission security§164.312(e)
- Transcription and AI involve no transmission at all. Optional sharing sends only a reviewed, identifier-redacted transcript.
Measured, not configured
How we chose the speech model
Care ships with one speech model, chosen for you: NVIDIA’s Parakeet TDT 0.6B v3, in FluidInference’s Core ML build, on iPhone and on Apple-silicon Macs alike. Core ML lets it ask for the Neural Engine. There is no model picker to get wrong.
We tested it the way a clinic would hear it: 57 mock consultations between clinicians and actors (PriMock57), plus a synthetic clinical corpus dense with drug names, doses and negation, against eight alternatives on the consultations, from Whisper tiny to WhisperKit’s large-v3-turbo, and one more, sherpa-onnx’s Parakeet v2, on the synthetic material only. Parakeet v3 was the most accurate engine among those run on every consultation, at a fifth of the memory of the GGUF build the app ran before (255 against 1,210 MiB).
Every engine we tried stumbled on medication names, and Care doesn’t pretend otherwise. A correction layer checks the transcript against RxNorm’s list of prescribable names and shows each change on a copy of the transcript: underlined, explained, one tap from reverting. Then a medication-verification step asks you to confirm every drug before a note is final.
Before anything can leave the device, an identifier model and a set of rules flag names, dates, numbers and places for your review, span by span. The review is the safeguard; the model only feeds it. Audio stays on the device.
- 10.8 %mean word error rate on the 57 PriMock57 mock consultations: ahead of Whisper tiny, base and small by mean over the full 57-file set, not on every file.
- 0.0 ptword-error cost of the drug-name correction layer on PriMock57, for +14.3 points of medication-term recall on synthetic dictation (tuned in-sample).
- 70.8 %of patient-identifier spans the flagging model found on its own in a 30-case synthetic corpus (46 of 65), before the rules and before your review.
- 255 MiBpeak memory on the 57 consultations on the test Mac; 299 MiB, and flat, on a 28.5-minute recording.
- Numbers from one Mac on October 11, 2026. No iPhone has been measured yet; the report names the rules that decide with device numbers. The correction layer, the verification step and identifier flagging are designed for this version and not yet measured end to end; the 0.0-point and 70.8 % figures are the lexicon pass’s and the identifier model’s on their own.
How it works
Online once.
Private every visit after.
Local AI does the typing. You check the source, verify the medications and decide when a note is final.
- 01
Set up once, online
Sign in with your Scribed account. Care downloads and verifies its models, then creates your encrypted vault. After setup, you can work with the connection off.
- 02
Record in the room
Confirm the patient and their permission to record. Audio is captured and transcribed on the device, in airplane mode if you like.
- 03
Review, verify, finalize
Correct the transcript, draft the note with the local assistant, confirm every medication and mark the note reviewed.
Does any patient data go to the cloud?
Not for transcription or AI. Audio, transcripts, notes and the assistant’s output stay on the device, with no cloud fallback. Two optional features are off by default: syncing your patient list to your own Scribed workspace, and sharing a transcript after you review it and remove identifiers. You choose when, if ever, to use them.
Is Scribed Care HIPAA compliant?
No app can be HIPAA compliant on its own. Compliance depends on your practice’s safeguards, policies and agreements, and we won’t put a badge on it. Care is designed so that transcription and AI never put patient data in a third party’s hands: it supports your access controls and audit trail with device authentication, encryption and a local audit log. Scribed does not yet offer a business associate agreement, so cloud sync stays unavailable to covered entities.
What happens if a phone is lost?
The vault is encrypted with a key that lives only in that device’s keychain, is tied to its passcode, and is never included in iCloud or other backups. Without the device owner’s authentication, the data can’t be read. An encrypted, passphrase-protected vault archive you create is the only copy outside the device.
Can I use Scribed Care today?
Care is in development for iPhone and Apple-silicon Macs. Contact us about preview availability. The development preview is for synthetic test data only and is not ready for real patient use; device, clinical and operational validation remain incomplete.
Which devices are supported?
iPhone on iOS 17 or later with a Neural Engine and at least 6 GB of memory, and Apple-silicon Macs. Note drafting and Ask need at least 8 GB. Android and Intel Macs are not supported, because the speech engine runs on Apple’s Neural Engine.
What is the local AI intended to do?
Help organize encounter documentation and find information in a saved transcript, with the source available for review. It is not designed to diagnose, prescribe or take clinical actions.
Private by architecture
Look at your patient.
Not at a cloud contract.
Get in touch about the Scribed Care development preview.
Ask about the preview
iPhone & Apple-silicon Mac · In development
For synthetic test data only. Not ready for real patient use. Device, clinical, and operational validation remain incomplete.