Skip to main content
NuBench AI
Get started

Trust Center

How NuBench AI protects your data

The architecture, controls, and posture behind the platform — written for the people who run your security review.

The 16-step governance pipeline

Every request to every model passes through the same sixteen steps — there is no ungoverned side door. Governance is the request path, not middleware bolted to the side of it.

  1. 1Authenticate
  2. 2Resolve tenant
  3. 3Rate limit
  4. 4Budget pre-check
  5. 5Classify prompt
  6. 6PII / PHI scan
  7. 7DLP policy
  8. 8Injection screen
  9. 9Route to provider
  10. 10Budget deduction
  11. 11Dispatch to model
  12. 12Classify response
  13. 13Response DLP
  14. 14Budget reconcile
  15. 15Audit write
  16. 16Usage metrics

Strong isolation, dedicated deployment available

Tenant data is logically isolated by schema and key prefix — cross-tenant queries are not possible by design. Enterprise plans go further: dedicated single-tenant deployment with exclusive database, cache, buckets, and encryption keys, so noisy-neighbour risk is structural rather than probabilistic.

Serverless-first compute

Governance services run as AWS Lambda functions that scale with request volume and cost nothing at rest; the always-on surfaces (gateway, consoles, chat) run as ECS Fargate services behind a load balancer. Failed deployments roll back automatically via a deployment circuit breaker.

Managed, replicated state

Aurora Serverless PostgreSQL scales capacity with load and is reached only through RDS Proxy, which pools connections and rides through failovers. ElastiCache Redis backs rate-limiting, budget counters, and collaboration state, with transit encryption on.

Everything is code

The entire environment — networks, keys, services, policies — is Terraform, applied through CI with review gates and static security analysis. An environment can be reproduced from the repository, which is also what a credible disaster-recovery story is made of.

Want the operational numbers?

Capacity targets, performance characteristics, and availability commitments are part of the procurement conversation, where we can match them to your workload rather than publish a number without context. Your security and infrastructure teams can also request a technical deep-dive against the infrastructure code itself.