AI Privacy Gateway
Runs server-side, on the wire.
Self-host it in your own cloud or run it DataSitr-hosted in-Kingdom. The detector, policy engine, router, and three-lane routing detailed below are this path.
01Deployment
Deployment.
Choose where DataSitr runs. The components stay the same, so your evidence and review exports stay consistent.
02What you deploy
One layer. Three products.
Deploy one, two, or all three — reduce the risk you can, contain the risk you can't. Each part runs in a different place; the architecture, runtime, and infrastructure detailed below describe the Gateway.
On the roadmap
In development — not offered yet.
Data Sitr Shield (on-device DLP) and Adversarial Prompt Protection (prompt-injection defense) are in active development. They are not yet available to deploy or pilot, and we make no claim that they are. We'll list them here, with their limits in writing, once they're ready. Today, the AI Privacy Gateway is what we offer.
// The AI Privacy Gateway is the control you can verify today. Data Sitr Shield and Adversarial Prompt Protection are in development and not yet offered. We'll never blur that line.
03Deployment models
Three paths, same components.
Pick the one that fits your operating model — we keep the runtime identical so evidence and exports look the same to a reviewer regardless of who hosts.
DataSitr-hosted pilot
We provision and manage a Saudi-hosted environment with shared-state services where required. Fastest path to a working pilot.
1–2 days
Your Saudi-hosted cloud
Your team deploys on your Saudi-hosted infrastructure. Two paths supported: ACK guided Kubernetes, or single-VPS Docker Compose.
1–3 days
Your datacenter
Same components on your hardware. No cloud dependency. Operator owns network egress, backup target, and identity provider.
1–3 days
04Architecture
One path. Every request.
Every request follows the same path: TLS edge, then the API runtime that detects personal data, applies your policy, and chooses the allowed route. State, audit, and signed evidence live alongside the API on Saudi infrastructure.
Request flow · Saudi-hosted runtime · 2026 baseline
The runtime is identical across hosted-pilot, customer-cloud, and on-premises models. Helm guards 2× replicas with shared Postgres + Redis when high-availability mode is on; a single-VPS Docker Compose path is supported for early pilots that don't need it.
05Three-lane routing in detail
Each lane has explicit rules.
The router's lane decision determines what the downstream AI provider sees. Each lane has explicit rules, an explicit destination, and an explicit audit trail.
GREEN
Tokenized · external AI
PII is replaced with typed placeholders before the request leaves the gateway. The external model never sees real identifiers, dates, or location names — only shape-preserving tokens.
AMBER
Pseudonymized · in-Kingdom AI
Linkable pseudonyms substitute identifiers but the prompt stays inside Saudi infrastructure. Use this when the workload needs internal context the green lane would strip.
RED
Raw · in-Kingdom only or blocked
Sensitive categories (health, national ID combinations, financial accounts) stay raw and route only to operator-configured in-Kingdom paths. If no in-Kingdom path is configured, the request is blocked.
06What gets deployed
The same components, every model.
The same five components run in every deployment model. Sizes scale with traffic; the topology stays the same.
datasitr-api
FastAPI · Python 3.12 · 2× replicas with shared state. Hosts the detector, policy engine, router, and audit writer.
dashboard SPA
React + Vite · pre-built static bundle served by the API container. Operator UI, regulator portal, academy.
postgres
Operational data store · sessions, attempts, audit, hash-chained compliance records. Daily backup with restore drill.
redis
Coordination layer · shared cache for rate-limits, OIDC sessions, and cross-replica state. Stateless after restart.
object storage
Encrypted backups + signed evidence outbox. Saudi-hosted bucket; off-host copy for regulator-readable export packages.
nginx (TLS)
Edge layer · TLS termination, HSTS, request size cap, health probes. Let's Encrypt or operator-supplied certificate.
07Runtime options
Pick the runtime that fits the host.
Runtime
When to choose
Notes
Alibaba ACK
Kubernetes · production
Kubernetes · production
Production deployments needing horizontal scaling, rolling updates, and managed control plane.
Guided deploy script · helm chart · 2× replicas with shared Postgres + Redis · health probes and recovery gates.
Docker + Compose
single VPS · pilot
single VPS · pilot
Early pilots and on-premises installs that don't need horizontal scaling yet.
Single command, dashboard build, health checks. SQLite-backed sessions when Postgres is not yet provisioned.
systemd
bare host · constrained
bare host · constrained
Environments where Docker is not available (regulated estates, air-gapped reviewers).
Requires Python 3.12+ and Node.js 20+ on the host. Operator owns process supervision and log rotation.
08Authentication
Keys for machines. SSO for people.
Method
Detail
API key
default · machine actors
default · machine actors
Bearer token in the Authorization header. Keys carry the sv_ prefix and are role-scoped (tenant / tenant_admin / super_admin / regulator).
OIDC SSO
optional · human actors
optional · human actors
Authorization Code + PKCE flow against the operator's corporate IdP. Person-bound identity for every action — required for individual training records and per-user audit attribution.
09Infrastructure requirements
Small for a pilot. Scales for production.
Minimum (pilot)
Recommended (production)
2 vCPU
4 GB RAM
20 GB SSD
Ubuntu 22.04+
Saudi region
single-VPS Docker Compose path
4+ vCPU per node · 2 nodes
8+ GB RAM per node
50+ GB SSD per node
Shared Postgres + Redis
Off-host backup + scheduled restore drill
Active alert delivery to operator on-call
10What gets verified
The runtime emits dated, hash-chained, regulator-readable evidence.
Operator-refreshed controls — not timeless freshness guarantees.
- Signed export packages Time-bounded, scope-bounded JSON manifests with key fingerprint and hash-chain verification — what a regulator opens first.
- Audit trail Every lane decision, every rights-request action, every key invocation — recorded once with a route reason, time, and actor identity.
- Training records The dashboard's built-in academy emits dated, exportable training records by named actor (when OIDC) or by shared key — distinguished and attributable.
- RoPA + transfer register Article 31 / Article 29 anchor surfaces with dated entries that link each transfer to its lawful basis and TRA reference.
- Subject-rights queue Verification, fulfilment, and dated closure evidence linked to a single workflow — auditable end-to-end without leaving the platform.
- Breach register PDPL Implementing Regulation Article 24 fields — detection time, awareness time, affected categories, supervisory-authority notification timestamp — with a 72-hour clock from awareness, per PDPL Article 20.
11Shield deployment
Data Sitr Shield — in development.
Data Sitr Shield (on-device DLP for the browser / GenAI channel) is in active development and not yet available to deploy or pilot. When it ships it is designed to install through your existing MDM and run locally on the device — nothing from DataSitr in the request path. We'll publish the deployment guide here once it's ready, with its limits in writing.
12What this is — and isn't
What this is — and isn't.
Current pilot runtime uses a Saudi-hosted shared-state layout. Dated proof covers scaling beyond a single-process setup, and dated alert-delivery and backup-plus-restore evidence is operator-refreshed. Treat those as controls the operator maintains, not as timeless freshness guarantees.
Optional immutable-evidence retention can strengthen audit evidence when configured. That is a software-level control — it should not be read as hardware-backed immutability or as a high-availability claim.
DataSitr is registered with NDGP as a data services / products provider (LR-25-000018, status Complete) — not licensed. We hold no SDAIA accreditation (the AE-26-000237 application was not awarded). References to PDPL articles describe operational alignment; they do not imply regulator approval.
See it work on your data.
Same components, hosted in the Kingdom — pick where they run and the evidence travels with you.