Skip to main content
When someone calls your Rivet number and leaves a voicemail, the audio follows a short, well-defined path — and the most sensitive parts of that path stay on Canadian hardware. This page is specific about where each step runs. It’s the level of detail your College, your compliance officer, or an HIC compliance review will ask for.

​
The path a voicemail takes

  1. Your client calls your Rivet number.
  2. Twilio (United States) records the audio. Twilio retains the recording for 30 days by default.
  3. Rivet’s worker — running in Canada — fetches the audio from Twilio for transcription.
  4. Whisper, an open-source speech-to-text model, transcribes the audio. Whisper runs locally on Rivet’s Canadian hardware. The audio file is read into memory, transcribed, and never transmitted to any cloud AI service.
  5. Ollama, running the same locally on the same Canadian hardware, classifies the transcript’s intent (new client, reschedule, urgent, etc.) using an open-source language model.
  6. The transcript and the intent label are written to Rivet’s database.
  7. You receive a push notification, see the voicemail in your inbox, and the appropriate auto-reply (if any) is sent to the caller.

​
What runs locally in Canada

  • Voicemail audio transcription — Whisper, on Rivet’s hardware in Belle River, Ontario.
  • Intent classification — Ollama, on the same hardware.
  • The voicemail-processing worker — also on the same hardware.
That hardware never sends audio or transcripts to OpenAI, Anthropic, Google, or any other cloud AI provider. The transcription and classification models are open-source and run as local processes. No third-party AI service ever sees a client’s voicemail.

​
What still touches the United States

Being specific about what is in the US matters as much as being specific about what isn’t.
  • Twilio records the inbound audio and holds it for the platform retention window (30 days by default). Twilio operates under its own published data-handling commitments and processes the audio only to deliver the recording back to Rivet.
  • The database (Supabase) — where the transcript and metadata land after processing — is currently hosted in a United States region.
  • Mobile push notifications route through Apple (iOS) and Google (Android FCM) push services in the United States. The notification payload doesn’t contain transcript text — it carries an identifier that your device uses to fetch the full content over HTTPS.
The Data Processing Agreement enumerates each sub-processor and its processing location. See the DPA for the full table.

​
Why the “no cloud AI” line matters

The most sensitive moment in a voicemail’s life is its raw audio. A caller may say their full name, the medication they take, a description of their symptoms, or the name of a family member. That’s the kind of content that, once it leaves your boundary into a cloud AI vendor’s training pipeline, becomes very hard to track. Running transcription and classification locally on Canadian hardware means:
  • The audio is read, transcribed, and discarded from memory.
  • The transcript is the only artifact that touches the database.
  • No vendor’s terms of service govern what happens to the audio in transit, in training, or in their analytics.
  • A subpoena to an AI vendor won’t pull a copy of your client’s voicemail, because the vendor never had it.

​
What this lets you say

If you’re describing Rivet’s privacy posture in a sales conversation, to your College, or to an HIC compliance review, the accurate phrasing is:
Voicemail audio and transcription are processed locally on Canadian hardware. No third-party AI service receives client audio or transcripts. The database is currently hosted in a United States region.
Avoid the broader phrase “Canadian data residency” — that overclaims on the database. The locally-run AI pipeline is the substantive Canadian processing claim Rivet can make today.

Client data handling

The full retention schedule for transcripts, audio, and conversation history.

Encryption

How the audio and transcript are protected in transit and at rest.

The Data Processing Agreement

The full sub-processor table and their processing locations.