API content
Audio and model outputs are processed only to return the requested result and are not retained after the response.
The public controls below describe the production Speech API and developer portal. They deliberately avoid claiming certifications or contractual service levels oruk does not currently hold.
Last updated August 5, 2026
Audio and model outputs are processed only to return the requested result and are not retained after the response.
Customer audio and outputs are not used to train, fine-tune, or evaluate models unless you explicitly agree otherwise in writing.
Passwords and API key secrets are stored only in non-recoverable form. Complete API keys are shown once.
Public website, account, billing, and inference traffic use encrypted HTTPS connections.
Account, billing, audit, and usage metadata use provider-managed encryption at rest and role-restricted access.
Card details are entered on Stripe-hosted pages and are not stored by oruk. Verified webhooks are required before credits are issued.
This is the default, and it is structural rather than only contractual: the usage ledger has no column able to hold content, and no results table exists in the schema. Audio and outputs are not stored or used to train, fine-tune, or evaluate any oruk model unless you explicitly agree otherwise in writing.
| Data | Retention |
|---|---|
| Audio sent to the API | Not retained. Held in memory to produce the response, then discarded. |
| Model outputs — transcripts, emotion and speaking-style labels | Not stored. Returned to you and discarded. |
| Request metadata — request and key IDs, model, task, audio duration, cost, timestamps | Retained for billing, abuse prevention, and audit. |
| Account records — name, email, password hash | Retained while the account is open; deleted on request. |
| Billing and credit ledger | Retained as required for accounting and tax. |
A request reaches the Cloudflare edge at speech-api.oruk.ai, which terminates TLS and forwards it to the inference origin through a private Cloudflare Tunnel. The edge injects a shared secret, strips any client-supplied copy of it, and removes forwarded-IP headers before the call; a request that arrives at the origin without that secret is refused. Only billing metadata continues to the control plane.
01 / Authenticate
Keys are validated against active organizations, expiration, revocation state, and available prepaid balance.
02 / Reserve
Each request reserves its maximum billable amount before inference, preventing concurrent overspend.
03 / Settle
Successful inference appends one idempotent usage debit. Failed requests release their reservation.
No payment provider or marketing analytics service receives customer API audio.
| Provider | Purpose | Processing location |
|---|---|---|
| Cloudflare | Website delivery, account control plane, usage records, edge security, and transactional email | Global routing; account records in managed infrastructure |
| Google Cloud | Speech inference | United States |
| Stripe | Payment processing, saved payment methods, receipts, and invoices | Stripe-managed regions |
| Resend | Fallback transactional email delivery when configured | Provider-managed regions |
| Google Analytics | Public marketing-site analytics only; excluded from authenticated account pages | Provider-managed regions |
Better learned here than in a questionnaire. These are the gaps a security or clinical-safety reviewer would otherwise have to find on their own.
No SOC 2, no ISO 27001, and no third-party security audit.
No HIPAA Business Associate Agreement. Protected health information should not be sent to the public API. Raise it with us first if PHI is in scope.
No multi-factor authentication on portal accounts yet. Google sign-in is available.
No published DPA. Data-processing terms are handled during enterprise review.
Retained metadata has no automated expiry. Access, correction, and deletion requests are handled by a person at access@oruk.ai.
No contractual SLA on pay-as-you-go. The status page reports reachability, not a commitment.
Model accuracy, demographic fairness, and the limits of emotion and speaking-style labels are covered separately under responsible use, with benchmark methodology and caveats on the methodology page.
Report a suspected security issue to access@oruk.ai. Reports are triaged by impact, contained, investigated, and communicated to affected customers when required.
Security questionnaires, data-processing terms, and deployment requirements are handled during an enterprise review. Public commitments remain in the Privacy Policy and Terms.