Security
Where your data goes, and who can see it.
Every claim below maps to a control we can point at, not a line of marketing. It also says what we don't claim, because a security page that only ever says yes is not one.
Every claim, its mechanism, how to verify it
| Claim | Mechanism | Verify |
|---|---|---|
| Your data is isolated at the database level. A query without your tenant context returns zero rows, never another customer's data. | Postgres row-level security, enforced on every tenant table, checked on every startup. | An automated posture check runs on every deploy. |
| Coverage isn't a hand-kept list. Every table holding tenant data is isolated automatically, derived from the schema itself. | Isolation is derived from model metadata, not a maintained list; a small reviewed exception covers only internal ops-telemetry tables that never carry customer rows. | Reviewed alongside the isolation policy itself. |
| Every API call needs a per-tenant key. Anything public is a reviewed, named exception, not an oversight. | Hashed per-tenant keys; authentication is on by default everywhere else. | The public-endpoint list is a standing internal audit. |
| No credentials or API keys live in our source code. | Secrets in environment variables only. | A pre-commit and CI scan blocks any that slip in. |
| An ERP's emailed export is only accepted from a sender we've verified for that workspace. | SPF or aligned DKIM checked against the domain that first delivered for you. | Enforced before a single row is written, on every inbound email. |
| No customer personal data sits in our application logs. | Log-scrub conventions; aggregate counts and ids appear, customer data does not. | Checked in periodic internal audits. |
| The AI cannot invent a number into your briefing. | A guard reads every AI answer before you do and discards any money figure no verified tool produced. | Runs on every AI answer, not sampled. |
| The AI only ever sees your own tenant's data, fenced off from instructions, never another customer's rows, your raw credentials, or your ERP password. | Tool results are tenant-scoped and fenced as data; ERP credentials are encrypted and never enter a prompt. | We don't claim your data never leaves us. Tool results do reach the AI provider, minimal and fenced. |
| Nothing the AI proposes executes without you approving it first. | Every write is a two-step confirm, audited and role-gated. There is no autonomous-write mode. | Every proposed action is a draft you click to send. |
| Your card never touches Narada's servers. | Checkout runs through Dodo Payments as merchant of record; Dodo, not Narada, holds PCI DSS Level 1 status. | We don't claim Narada itself is PCI-certified, we're out of scope by design. |
| Narada never writes to your ERP. Read-only, always. | Every connector call site is a read; none issues a write. | We don't claim a technical block would reject a write call if one were ever added, only that none exists today. |
| Ask, and we export everything you ever gave us. | A full export of every tenant table, run by us on your request. | We don't claim this is a self-serve button. An operator runs it for you, not you for yourself. |
| Deleted means deleted. Ask, we confirm, it's gone. | Every tenant-scoped row is hard-deleted; the operator must echo your exact workspace name before it runs. | Irreversible once confirmed, and logged internally. |
For your IT and finance gatekeepers.
The legal side, data-processing terms, encryption, deletion, export, lives on one page, written to be forwarded.