Platform
Architecture
Last updated March 3, 2026
Understand the fundamental concepts behind the Cencori platform: CIP, Data Residency, and Global Infrastructure.
Cloud Intelligence Provider (CIP)
Cencori is not just an "AI Proxy." We are a Cloud Intelligence Provider (CIP).
- AI Proxy: A simple pass-through layer that forwards requests. (e.g., Cloudflare AI Gateway).
- CIP: A fully managed infrastructure layer that provides stateful intelligence. Cencori maintains context (Memory), enforces security (Scan), and orchestrates multi-step logic (Workflows) before requests ever reach a model provider.
Request Lifecycle
Understanding the flow of a single request helps clarify where value is added:
- Application: Your code calls the Cencori SDK.
- Edge Network: Request hits the nearest Cencori Edge node (deployed globally across 35+ regions).
- Security Scan: Payload is analyzed for PII, prompt injection, and policy violations. Latency: <10ms.
- Routing Engine: Cencori determines the optimal provider based on cost, status, and your project rules.
- Provider Execution: Request is forwarded to OpenAI/Anthropic/Google.
- Observability: Response is logged, token usage is calculated, and async evals are triggered.
- Response: Final payload is returned to your application.
Cold Starts: Cencori's edge network is built on warm execution environments. There is zero cold-start latency for standard API requests.
Data Residency & Privacy
We take data sovereignty seriously.
- Logs & Analytics: Stored in our secure US-East (N. Virginia) cluster by default. Enterprise plans support EU-only storage.
- AI Memory (Vector Store): Embeddings and vector data are encrypted at rest in dedicated isolated clusters.
- Ephemeral Processing: Determining PII or routing logic happens in-memory on edge nodes and is never persisted to disk.
Global Infrastructure
Cencori is multi-cloud and multi-region by default.
- Availability: 99.99% uptime SLA for Enterprise.
- Regions: We route traffic through AWS, GCP, and Cloudflare networks to ensure lowest-latency connectivity to AI providers.
- Failover: If a region goes down, traffic is automatically re-routed to the next nearest healthy node.
Platform Limits
To ensure fair usage and stability, the following limits apply:
| Limit | Free Tier | Pro Plan | Enterprise |
|---|---|---|---|
| Concurrent Requests | 10 | 100 | Unlimited |
| Requests per Month | 10k | 1M | Custom |
| Log Retention | 3 Days | 30 Days | 90+ Days |
| Team Members | 1 | 5 | Unlimited |
Keys & Authentication
- Project Keys (
csk_...): Scoped to a specific project. Used by your application SDKs. If compromised, only that single project is affected. - Organization Keys (Coming Soon): Admin-level keys for programmatic management of multiple projects.
Comparison
Why choose Cencori over other solutions?
| Feature | Cencori (CIP) | Helicone / Portkey | LangChain / LangSmith |
|---|---|---|---|
| Primary Goal | Production Infrastructure | Observability & Logging | Prototyping Framework |
| Security (PII/Injection) | Native / Built-in | Plugin / External | Manual Implementation |
| Unified SDK | Yes (One API) | No (Proxy only) | Yes (Abstractions) |
| Multi-Provider | Native | Native | Native |
| Vector Memory | Integrated | No | External (Vector DB) |
| Latency Impact | < 15ms | ~50-100ms | N/A (Client-side) |
Glossary
- Gateway: The unified entry point for all AI requests.
- Memory: Cencori's managed vector store for RAG and long-term history.
- Scan: The security engine that inspects requests for vulnerabilities.
- Namespace: A logical partition within Memory to isolate data (e.g., per-user or per-document).