Why we built Bryl

The 50-minute hour, the laptop on the kitchen table at 9 PM, and the simple thesis that the safest place for a therapy session is the iPhone in your pocket.

Every therapist I've talked to has the same Tuesday evening.

The last client of the day leaves at six. The kids are waiting. Dinner is on. They sit at the kitchen table, open the laptop, and start typing notes from memory — six of them, sometimes eight — until 8:30, sometimes 9. One of them will get forgotten. One detail will get half-remembered. The clinical note that gets stored in SimplePractice or TherapyNotes the next morning is a faithful approximation. It is not what happened in the room.

The cottage industry of AI scribes that has formed in the last two years exists to fix this. Most of them work. Most of them have a problem.

The cloud problem

The standard pitch from cloud-based AI scribes goes something like: we're HIPAA-compliant, we have a BAA, your data is encrypted in transit and at rest. It's technically true, and it's the wrong frame.

Encrypted-in-transit means the audio is encrypted on its way to a server you don't own. Encrypted-at-rest means it's encrypted on a disk you don't own. A Business Associate Agreement is a legal contract that says if the vendor breaches your patients' data, the vendor is liable. It is a thin layer of paper that does not change the fact that an unencrypted copy of your client's session existed, briefly, on a machine you do not control.

The therapists I trust did not pick this profession because they enjoyed paperwork. They picked it because they have a high tolerance for sitting with another human's pain and a low tolerance for adding to it. Asking them to trust a cloud, the cloud's staff, the cloud's subprocessors, the cloud's breach response plan, and the cloud's SOC2 auditor — when the alternative is doing the work on the device in their hand — is the wrong ask.

The simpler answer

The iPhone in your pocket has more compute than the server farm that ran ChatGPT in 2022. The new generation of Apple Silicon, with its Neural Engine and unified memory, can run a 3.8-billion-parameter language model and a Parakeet-class transcription model entirely on-device, fast enough to draft a SOAP note before the client's shoes are off in the parking lot.

If the iPhone can do the work, the audio doesn't need to leave it.

That is the entire thesis of Bryl. We did not invent on-device transcription. We did not invent on-device language models. Apple shipped Foundation Models in iOS 26; the FluidAudio team ported NVIDIA's Parakeet TDT v3 to Apple Silicon. The thing nobody had done was wire those primitives together into a tool that respects how a 50-minute talk-therapy session actually unfolds — pauses, silences, two voices, six different note formats, a population that is structurally risk-averse — without ever requiring the practitioner to upload a single byte of session audio anywhere.

What we're building

Bryl is the on-device scribe for licensed psychotherapists. Three taps:

1. Capture. Open Bryl. Tap Record. The audio is protected by NSFileProtectionComplete from the moment of capture and sealed with AES-256-GCM the instant recording stops, keyed to a value held in the iOS Keychain.
2. Process. When you're done, Bryl transcribes the session locally with Parakeet TDT v3 (or Apple's built-in Speech framework as a fallback), replaces PII with stable placeholders before the language model sees the transcript, and drafts a SOAP, DAP, BIRP, EMDR, Intake, or Progress note using Apple Foundation Models or a bundled Phi-4-mini fallback.
3. Sign. Read the sections. Use Smart Edit to give the model a specific instruction (“shorter assessment”, “client's own words”, “CPT-aligned objective”) and rewrite. Press and hold for 1.5 seconds to seal — a SHA-256 hash and signature block are written into a notarial frame. The audio file is overwritten the moment the seal commits. The note is yours.

You copy the signed note into your EHR by hand. We deliberately do not offer an API integration that would require uploading PHI in transit to our servers. The friction of one paste is, in our view, a feature, not a bug.

What this is not

It is not a generic AI scribe with a privacy-flavored marketing skin. The architecture is the product.

It is not a therapy-tech app that wants to do everything. It does one thing — capture the session and draft the note — and it does that thing in a way that's falsifiable. Turn off your network. The app keeps working.

It is not a SaaS-startup pitch. There is no cloud infrastructure to amortize. The economics of a subscription business that has no cloud are different — and we think, ultimately, more sustainable — than the economics of a business that has to reduce its margin every time a model gets cheaper or a hyperscaler raises its prices.

A note on what we won't do

We will not build a Bryl cloud. If we ever do, this document and the architectural attestation we publish alongside it both stop being true, and we will tell you so before they stop being true. We will not introduce ad SDKs. We will not sell the marketing-site analytics to anyone. We will not, under any pressure, send your session audio to a third-party LLM provider in exchange for a slightly better note draft. The whole point of the architecture is that the question never arises.

If you are a licensed psychotherapist in private practice in the US — or somewhere that the App Store reaches — we're inviting 500 of you into the closed beta in Q3 2026. Tell us a little about your practice.

The audio belongs in the room.