Banking and capital markets
Account numbers, routing details, and customer identifiers are tokenized before any model call, with fail-closed enforcement when inspection cannot complete.
Data sovereignty · Regulated AI · Self hosted
Sovereign AI Privacy Gateway
A customer-hosted control point that inspects every prompt, file, and tool payload before it leaves the trust boundary — classifying regulated data, tokenizing or blocking it under policy, and writing encrypted audit evidence that regulators can actually read.
100%Containment across inspected traffic: every flagged payload was tokenized or blocked before it reached an external model.
15Regulated entity classes inspected per request, from PAN and CVV to SSN, IBAN, passport, and private keys.
2 msMedian inspection latency added to the request path, with a 21 ms p95 measured under enforcement.
4Shipped regulatory policy templates covering financial services, healthcare, cardholder data, and EU privacy.
System architecture
One inspection point between your people and every external model.
The gateway is a reverse proxy you run yourself. Prompts enter in plaintext, leave tokenized, and return rehydrated for the authorized user. Policy, token mappings, and audit evidence never cross the boundary.
Live control console
Evidence an auditor can read without reading your data.
Screens captured from a running deployment under live enforcement. Sensitive values appear only as masked previews and stable token references, never as recoverable content.
Where it fits
Built for the industries that cannot simply paste into a chatbot.
Every industry below maps to a policy template that ships with the gateway, so the enforcement contract starts from a recognised regulatory baseline rather than an empty configuration file.
Insurance
Claims, underwriting notes, and policyholder correspondence carry dense personal data. Detection runs on prompts, attached files, and tool payloads alike.
Healthcare and life sciences
Protected health information is classified and removed at the boundary, so clinical and research teams can use commercial models without exposing patient records.
Public sector and government
The entire inspection, tokenization, and evidence path runs on infrastructure the agency controls, which keeps residency and sovereignty commitments verifiable.
Payments and commerce
Primary account numbers are tokenized and sensitive authentication data is blocked outright, keeping model usage outside the cardholder data environment.
EU-regulated enterprises
Pseudonymization at the boundary supports data minimization arguments, and every decision is recorded as evidence rather than asserted in a policy document.
Compliance coverage
Regulatory intent translated into enforceable controls.
Each framework below is expressed as concrete gateway behaviour — what is detected, what happens to it, and what evidence survives. The emblems are original marks used to identify the framework, not certification seals.
United States · HealthcareHIPAA and HITECH
Protected health identifiers are detected and removed before disclosure, technical safeguards are enforced in the request path, and encrypted audit records support breach-notification analysis.
Global · Cardholder dataPCI DSS 4.0.1
Primary account numbers are tokenized rather than transmitted, sensitive authentication data such as CVV is blocked outright, and the audit trail never stores recoverable card content.
European Union · PrivacyGDPR
Pseudonymization at the boundary supports data-minimization and purpose-limitation obligations, and per-request evidence gives data protection officers a defensible processing record.
United States · FinancialGLBA Safeguards Rule
Non-public personal information is classified and contained before it reaches a third-party model, giving the safeguards programme a technical control rather than a written promise.
United States · Banking supervisionFFIEC guidance
Third-party model usage becomes an inspected, logged, and reversible flow, which is the evidence examiners expect for outsourced technology and vendor risk review.
European Union · ResilienceDORA
Fail-closed enforcement means an inspection failure stops the request instead of leaking data, and the encrypted evidence store supports ICT incident reporting timelines.
Compliance note: these are technical control mappings that support a compliance programme. They do not by themselves certify legal compliance with any regulation — formal certification remains the responsibility of the deploying organisation. Framework names belong to their respective owners; the emblems shown here are original marks drawn in this site's own style for identification only.
Technology foundation
A performance-critical data path with a boring, auditable control plane.
Inspection sits in the latency budget of every request, so it is written in Rust. Policy, evidence, and operations change slowly and need durability, so they live in a Go service backed by PostgreSQL.
Rust · Tokio · AxumInspection and proxy data path
GoControl plane and policy APIs
PostgreSQLVersioned policy and encrypted audit
React 19 · TypeScriptAuditor dashboard built with Vite
Kubernetes · KustomizeCustomer-hosted deployment overlays
nginx · AES-256-GCMAuth edge and envelope encryption
Solution fit
Containment at the boundary beats a contractual promise.
A no-training clause is an assurance you cannot verify and cannot show an examiner. The gateway makes the same guarantee structural: regulated content is detected, tokenized, and evidenced before a request is ever permitted to leave the network.
Regulated teams that still need frontier models
Sensitive entities are detected and tokenized at the boundary, so the request leaving your network carries no regulated content. Because tokenization is reversible inside the boundary, the authorized user still receives a complete answer.
Audits that need evidence, not assurances
Every decision writes encrypted, masked evidence that survives the configured retention window, so an examiner can be shown what was inspected, what was contained, and exactly when it happened.
Programmes that need a policy on day one
Versioned regulatory templates ship with the product, so the control set starts from a reviewed baseline instead of an empty configuration file and a professional-services engagement.
| Capability | Typical market approach | Sovereign AI Gateway |
|---|---|---|
| How data is protected | A contractual no-training, no-retention commitment from the model provider. | Regulated content is detected and tokenized before the request leaves the boundary, so there is nothing sensitive to retain. |
| Behaviour on failure | Inspection failures fail open so that traffic keeps flowing. | Fail-closed by design: an inspection failure stops the request instead of leaking data. |
| Usefulness after redaction | One-way redaction leaves the model reasoning about placeholder text. | Tokenization is reversible inside the boundary, so the authorized user receives a complete, usable answer. |
| Audit evidence | Plaintext request logs, where they are kept at all, with short retention. | Encrypted, masked evidence for every decision, retained for the configured window. |
| Getting to a policy | An empty configuration file plus a professional-services engagement. | Versioned regulatory templates covering the major regimes, shipped with the product. |
| Cost of adoption | Replace the coding assistants your developers have already chosen. | Sits in front of existing tools, so teams keep the assistants they use today. |
| Where it runs | Multi-tenant SaaS in the vendor’s chosen region. | Customer-hosted Kubernetes with Kustomize overlays, inside your own jurisdiction. |
Comparison note: the market column describes the default posture of provider-hosted assistants and one-way DLP redaction rather than any named vendor product. Control mappings support a compliance programme; they do not replace legal certification, which remains the responsibility of the deploying organisation.
Related systems
Metric note: containment, latency, entity coverage, and decision-mix figures are measured on a running deployment enforcing the financial-services policy template. They describe that configuration and workload, and will vary with policy, payload size, and infrastructure.