Data architecture
Tenant isolation
Every customer workspace is a distinct tenant with a UUID. Row-level security in PostgreSQL enforces isolation at the database layer: every tenant-scoped query runs inside a transaction that sets app.current_tenant_id, and RLS policies gate every table by that setting. Cross-tenant reads are impossible via the application role.
Encryption at rest
Sensitive data (OAuth tokens, API keys, Slack bot tokens, MCP server credentials) is envelope-encrypted. Each column is wrapped by a per-row data-encryption key (DEK) which is itself wrapped by a per-tenant key held in our key-management system. Enterprise tiers can bring their own KMS key so we cannot decrypt customer data without the customer's cooperation.
Encryption in transit
All customer-facing traffic uses TLS 1.3. Internal service-to- service traffic inside our Kubernetes cluster uses mTLS via Traefik's in-mesh certificate rotation. Traffic to upstream LLM providers uses their published TLS endpoints.
Audit chain
Every agent run is recorded to a per-tenant, append-only audit log capturing the prompt, tool calls, tool results, model responses, token usage, and cost. PII is redacted at write time via a regex-based redactor (gitleaks-style pattern set for credentials; a fuller NER-based redaction pass is on the compliance roadmap). Audit rows can be exported by admins at any time.
How LLM data flows
Managed LLM mode
On Basic through Scale tiers (Managed mode), Gauss holds the upstream LLM API keys. Messages you address to Gauss are sent to the selected provider (Anthropic, OpenAI, Google, OpenRouter) under our account. Provider terms explicitly forbid training on API inputs; we do not authorize any other data use.
Bring-your-own-key (BYOK) mode
Scale and Enterprise tiers can supply their own LLM provider API key. In this mode, LLM traffic flows directly under the customer's account; the LLM provider bills them, not us. We charge a per-token platform rate for the infrastructure that carries the traffic; no upstream provider relationship exists between us and the LLM in this configuration.
Dedicated deployment
Enterprise customers with the strictest sovereignty needs can run Gauss as a dedicated single-tenant deployment inside their own AWS or GCP account, with their own LLM keys — no customer data leaves their environment. We deploy and manage the stack via infrastructure-as-code; the customer keeps root access to the cloud account.
Access controls
Authentication
Dashboard access is via Sign-in-with-Slack or Sign-in-with- Google OIDC. We do not store passwords. Session tokens are stateless ED25519-signed JWTs (24-hour validity) with a short-lived refresh token (7 days) rotated on use.
Authorization
Three roles: owner, admin, viewer. Owners and admins can manage integrations, change models, invite members, and edit workspace instructions. Viewers can see their own runs but cannot modify tenant configuration.
Slack integration scope
Gauss requests only the Slack scopes required to operate:app_mentions:read, chat:write,im:*, channels:history, and OIDC scopes for sign-in. We do not request scopes that would let us post to arbitrary channels without your knowledge.
Infrastructure
- Region: AWS eu-west-2 (London). EU-only data residency by default.
- Compute: EKS Kubernetes cluster; roles split into isolated deployments (webhooks, management, worker, scheduler) with separate service accounts.
- Storage: RDS PostgreSQL 18 with pgvector, ClickHouse Cloud for analytics, Redis for hot metadata cache.
- Secrets: Kubernetes Secrets sealed via Pulumi + AWS KMS. No secrets in the codebase or CI logs.
Certifications and audits
- SOC 2 Type II — engagement in progress with a Big Four auditor. Report available on request under NDA once the observation window closes.
- GDPR / UK GDPR — data processing agreement available for enterprise customers.
- HIPAA — available on the Enterprise Dedicated tier with a signed BAA.
Reporting security issues
If you discover a security vulnerability, please email security@gaussintelligence.io with details. We aim to respond within one business day.
Please do not publicly disclose issues before we've had a reasonable window to remediate. We're happy to coordinate on disclosure timing and credit researchers who report responsibly.
What's on the roadmap
- NER-based PII redaction to complement the regex pass currently in production.
- SAML/SCIM for enterprise customers.
- Customer-managed KMS keys on Enterprise Dedicated and above.
- Multi-region deployment for latency-sensitive customers.
- FedRAMP / HITRUST for Custom-tier engagements as demand justifies.