Draft
Draft. This document is under legal review and is not yet in effect.
This is the canonical list of subprocessors Sayli uses to run the service. Each one sees only what its job requires. All of them are bound by terms that prohibit training models on customer data.
| Subprocessor | What it does | Region |
|---|---|---|
| Google Cloud Platform | Hosting, storage, databases, and background jobs. All customer content lives here. | United States (us-central1) |
| Deepgram | Speech-to-text. Receives call audio, returns transcripts. | United States |
| AssemblyAI | Speech-to-text fallback when Deepgram is unavailable. | United States |
| Cerebras | Language-model inference for answers and briefs. Receives redacted transcript excerpts and questions. | United States |
| DeepInfra | Language-model inference fallback when Cerebras is unavailable. | United States |
| OpenAI | Text embeddings only, for search. No chat traffic, no audio. | United States |
| Pinecone | Vector search over embedded content, scoped per organization. | United States |
| Resend | Transactional email: sign-in links, invites, briefs you send. | United States |
| Lemon Squeezy | Billing and merchant of record. Handles payment data; we never see full card numbers. | United States |
| PostHog | Product analytics. Feature-usage events, never call content. | European Union (EU cloud) |
| Sentry | Error reporting. Reports are scrubbed of transcript text, questions, answers, and tokens before send. | United States |
Changes to this list
We will give notice before adding or replacing a subprocessor. Customers with a signed DPA can object on reasonable data protection grounds under its subprocessor clause.
Questions
Write to privacy@sayli.ai.