Building a Privacy-First AI Platform: Sevrel's Architecture
CRE documents contain confidential rent figures, deal terms under negotiation, and proprietary valuations. Sevrel is architected so your data is never used to train AI models, never leaks across organizations, and is always traceable.
The Architecture at a Glance
Sevrel's privacy architecture has three key components:
Enterprise AI Provider (Tiered Routing)
Sevrel uses an enterprise AI provider with tiered model routing (fast, standard, and deep tiers). All inference requests go out over TLS. Our AI provider does not use Sevrel customer data to train its models, and prompts are not retained beyond the transient processing window.
Per-Organization Scoping
Every query is scoped by organization ID before a single byte leaves the backend. PostgreSQL Row-Level Security policies enforce isolation at the database level — defense-in-depth that prevents cross-tenant data access even if application logic is bypassed.
Application & API Layer
The web application and API run on managed infrastructure with DDoS protection, TLS encryption, and edge caching. Session cookies are HttpOnly, Secure, and SameSite=Lax. No secrets or tokens are ever exposed to client-side JavaScript.
Privacy Design Principles
- No secrets in the frontend. API keys, tokens, and credentials exist only server-side. The browser never sees them.
- Session cookies only. HttpOnly, Secure, SameSite=Lax. No tokens in localStorage or sessionStorage.
- Encrypted credentials at rest. OAuth tokens for Egnyte and Microsoft Graph are encrypted with symmetric keys before database storage.
- Per-organization isolation. Every database query is scoped by tenant. No cross-organization data paths exist.
- Append-only audit log. Every document access, query, and admin action is recorded with correlation IDs for end-to-end traceability.
Tiered Model Routing
No single model is best at everything. Sevrel automatically routes each query across three tiers based on complexity:
Fast
Fastest responses. Used for titles, intent classification, and simple lookups.
Standard
Best balance of speed and depth. The default for RAG queries, agent workflows, and general chat.
Deep
Deepest reasoning. Reserved for deal underwriting, complex legal analysis, and systematic research.
What This Means in Practice
For a CRE organization evaluating AI tools, Sevrel's architecture answers the questions that IT security and legal teams care about most:
- “Where does our data go?” — TLS-encrypted to our enterprise AI provider, then discarded
- “Is our data used for training?” — No. Our AI provider contractually does not train on Sevrel customer data
- “Who has access?” — Only authenticated users within your organization; enforced by per-org scoping and Postgres RLS
- “Can we audit usage?” — Yes, comprehensive append-only audit trail with correlation IDs