Deployment models
| Model | Best fit | Customer control | Clarivy responsibility |
|---|---|---|---|
| Clarivy SaaS | Fast design-partner pilot or shared managed service. | Tenant policy, users, approved connectors and data requirements. | Application operations, shared platform controls and support. |
| Private cloud | Regulated customer requiring a dedicated environment. | Subscription, network, identity, keys and regional policy. | Product release, configuration, integration and agreed operations. |
| On-premises | Customer requiring local infrastructure and restricted egress. | Compute, network, storage, identity, backup and runtime access. | Licensed software, deployment package, documentation and support boundary. |
Azure reference architecture
The recommended production path uses managed services rather than running every dependency on one host. Container Apps scale the web, API and workers independently; PostgreSQL remains authoritative; Blob Storage holds evidence artifacts; Service Bus carries retryable work.
Production request path
- 1
Controlled edge
Staff, merchants and partner systems enter through managed TLS, web-application firewall and API controls.
- 2
Identity and authority
Enterprise identity proves who the caller is; Clarivy computes tenant, role, workflow and segregation-of-duties permissions.
- 3
Short case transaction
The API records authoritative state and sends expensive or retryable work to a queue.
- 4
Independent workers
Evidence, document and communication workers scale from queue depth with bounded retries and idempotency.
- 5
Controlled intelligence
OCR and model providers return provisional structured output through the existing Clarivy harness.
- 6
Observable outcome
Metrics, traces, audit lineage and cost-per-case remain available without logging sensitive document or prompt content.
Scaling stages
| Stage | Runtime pattern | Reliability and cost posture |
|---|---|---|
| Design-partner pilot | One managed environment, one warm API, workers scale to zero, PostgreSQL with vector extension. | Backup-based recovery and strict budget caps. |
| Controlled production | Two warm API replicas, queue-scaled workers, private data services and managed edge protection. | Zone-redundant database, formal alerts and restore exercises. |
| Regulated enterprise | Dedicated or customer-hosted environment with customer-specific network and key controls. | Contracted recovery, isolation, retention and operational evidence. |
Acceptance gates
- Every customer-owned query is tenant constrained.
- Client input cannot choose a storage path or audit identity.
- Policy versions, hashes and model runs remain reconstructable.
- Duplicate events do not create duplicate effects.
- AI and OCR outages degrade safely to manual review.
- Restore objectives are agreed and proven.
- Secrets are managed outside repositories and browser code.
- Residency, retention and provider boundaries are contractually approved.
Architecture guidance, not a certification
Continue reading
Technical and compliance glossary
Review the terms used across the production architecture and control model.