VDF AI is an on-prem AI agent platform for regulated enterprises that need air-gapped AI deployment, private RAG, sovereign AI infrastructure, and customer-managed everything. Nothing depends on an external inference API. This is where security, compliance, and procurement teams find how it works.
These are the questions buyers, auditors, and AI answer engines ask first when comparing cloud AI, on-prem AI agents, sovereign AI infrastructure, and air-gapped deployments.
Run orchestration, private RAG, embeddings, model inference, logging, and governance inside your own perimeter. VDF AI is designed so prompts, documents, embeddings, model outputs, and audit logs stay on-premises, in your private cloud, or in an air-gapped environment.
It means the production data plane has no required path to third-party model APIs. Network egress can be disabled, models can run locally, and security teams keep control of identity, keys, retrieval indexes, logs, and update approval.
Private RAG grounds answers in your own documents and databases while keeping retrieval, vector indexes, embeddings, source permissions, and citations inside your controlled infrastructure.
Every model runs inside a boundary you control. Choose the isolation level your regulators, security team, and infrastructure require — the platform is the same.
The full platform — orchestration, routing, retrieval, and models — runs on networks with no outbound internet access. Updates arrive through a controlled, offline artifact process.
Deployed on your own hardware in your data center. Data, models, and logs remain inside your infrastructure and identity boundary, governed by your existing controls.
Deployed into your own cloud tenancy (AWS, Azure, GCP, or sovereign cloud). You keep the account, the network policy, the encryption keys, and the data plane.
VDF AI plugs into the identity, logging, secrets, and network controls you already run — it does not ask you to trust a new external service.
Integrates with your existing identity provider over SAML or OIDC (Entra ID, Okta, Keycloak, LDAP). No separate user store to manage, no shadow accounts.
Granular RBAC governs who can use which agents, models, tools, and data sources. Access to sensitive workflows and connectors is scoped by role, team, and environment.
Every prompt, tool call, retrieval, model route, and response is recorded with actor, timestamp, and context. Logs stream to your SIEM for retention and investigation.
You control which models are approved, where they run, and which workloads may use them. The router enforces model policy per domain, sensitivity level, and residency requirement.
Retrieval runs against your own document stores and vector indexes inside your perimeter. Embeddings are generated locally; no content is sent to third-party APIs.
Deploys into segmented network zones with no dependency on external inference endpoints. Egress can be fully disabled for air-gapped and OT environments.
Policies and contracts are necessary, but regulated AI deployments also need enforceable technical boundaries. The architecture should make the safe path the default path.
| Decision area | Typical cloud AI platform | VDF AI private deployment |
|---|---|---|
| Inference path | Prompts and context are commonly sent to an external model endpoint. | Models can run inside the customer perimeter; external inference is optional and can be disabled. |
| Retrieval and embeddings | Documents and embeddings may be processed or stored in provider-managed services. | Private RAG, embeddings, indexes, and citations stay in customer-managed infrastructure. |
| Data residency | Residency depends on provider regions, subprocessors, service settings, and contractual controls. | Data plane is pinned to the customer data center, private cloud tenancy, sovereign cloud, or air-gapped network. |
| Audit evidence | Audit detail is limited by the provider’s platform events and retention model. | Prompt, retrieval, model route, tool call, response, and approval events stream to your SIEM. |
| Change control | Platform, model, and policy changes may follow the vendor’s release cadence. | Customer-approved artifacts, model versions, and updates support regulated change windows. |
| Cost model at scale | Per-token, per-run, or consumption meters can make broad agent adoption hard to forecast. | Platform pricing and customer-owned capacity support predictable scaling across teams. |
Contractual promises matter, but regulated buyers also need a technical boundary that prevents data from leaving in the first place.
Audit logs, model routes, retrieval citations, and approval records should be captured automatically as work happens.
Identity, keys, models, network policy, data stores, and retention should be governed by the enterprise operating the system.
On-prem deployment is the most direct way to satisfy data-residency, sovereignty, and third-party-risk obligations. Here is how the architecture supports each framework.
| Framework | Compliance support | |
|---|---|---|
| EU AI Act | Risk classification, human oversight, traceability, and technical documentation for governed AI workflows. VDF AI supports EU AI Act compliance programs with AI inventory, use-case risk classification, human-oversight controls, change history, audit logs, and technical documentation generated from governed workflows. | Details |
| GDPR | Data residency, purpose limitation, minimization, and deletion workflows for personal data. Data stays in-region and inside your perimeter; no production prompt or document processing is required by external model providers. Customer-managed retrieval and storage support minimization, purpose limitation, retention, and erasure workflows. | Details |
| DORA | Operational resilience evidence for ICT risk, incident review, testing, and third-party dependency reduction. On-premises and air-gapped deployment reduce reliance on external inference providers as operational dependencies. Full audit trails, model governance, offline update controls, and SIEM integration support resilience testing and incident reporting programs. | Details |
| HIPAA | PHI containment, access control, audit logging, authentication, and encryption support. For healthcare deployments, PHI can remain inside the covered entity’s environment. Customer-managed access controls, audit logs, authentication, network segmentation, and encryption support Security Rule-aligned AI-assisted clinical and administrative workflows. | Details |
| NIS2 | Cybersecurity risk-management support for essential and important entities. Air-gapped, on-premises, and network-isolated deployment options help essential and important entities keep critical AI workflows under local control, with logging, access controls, incident investigation data, and supply-chain visibility. | Details |
VDF AI supports EU AI Act compliance programs with AI inventory, use-case risk classification, human-oversight controls, change history, audit logs, and technical documentation generated from governed workflows.
Data stays in-region and inside your perimeter; no production prompt or document processing is required by external model providers. Customer-managed retrieval and storage support minimization, purpose limitation, retention, and erasure workflows.
On-premises and air-gapped deployment reduce reliance on external inference providers as operational dependencies. Full audit trails, model governance, offline update controls, and SIEM integration support resilience testing and incident reporting programs.
For healthcare deployments, PHI can remain inside the covered entity’s environment. Customer-managed access controls, audit logs, authentication, network segmentation, and encryption support Security Rule-aligned AI-assisted clinical and administrative workflows.
Air-gapped, on-premises, and network-isolated deployment options help essential and important entities keep critical AI workflows under local control, with logging, access controls, incident investigation data, and supply-chain visibility.
This mapping describes how the platform’s architecture and controls support your compliance program. It is not legal advice; obligations depend on your deployment, jurisdiction, and use case.
No. VDF AI is designed to run entirely inside your infrastructure — on-premises, in your private cloud tenancy, or fully air-gapped. Prompts, documents, embeddings, model weights, and logs remain inside your perimeter. There is no requirement to call an external inference API.
Yes. VDF AI is model-agnostic and supports bring-your-own-models. You choose which open-weight or licensed models to run, where they run, and which workloads may use them. The router enforces your model policy per domain, sensitivity, and residency requirement.
Authentication uses your existing SSO over SAML or OIDC. Authorization is governed by role-based access control. Audit events stream to your SIEM. Secrets integrate with your existing vault, and encryption keys can be customer-managed.
We provide a reference architecture, an RFP/evaluation checklist, deployment and network diagrams, a compliance mapping, and a security questionnaire response. We can also join a security architecture review with your team.
Yes. The complete platform — orchestration, routing, retrieval, and models — can run with no outbound internet access. Updates and models are delivered through a controlled offline artifact process suited to defense, government, and OT networks.
Deploy the full data plane inside your own environment: identity, orchestration, private RAG, vector indexes, embeddings, model runtime, audit logs, and admin controls. In this pattern, external inference APIs are not required, and egress can be disabled for sensitive networks.
Updates are packaged as signed artifacts that your team can inspect, approve, transfer through your existing offline media process, and install from an internal registry or artifact repository. The air-gapped runtime does not need outbound internet access.
Model governance is enforced through an approved model catalog, version history, policy-based routing, workload permissions, evaluation records, and audit logs showing which model handled each request. You decide which models are allowed for each domain and sensitivity level.
Yes. Audit events can be exported to your SIEM, routed into incident-response workflows, and used as evidence for GRC processes. Typical events include actor, agent, data source, model route, tool call, response, approval, and timestamp.
Documents remain in your approved stores or are indexed into your private vector database. Embeddings are generated locally, retrieval respects document permissions, and responses cite source material without sending content to third-party model providers.
Yes. VDF AI is designed for BYOM enterprise AI: bring your own models, embedding models, serving runtime, GPU infrastructure, identity provider, keys, and monitoring stack. The platform governs the workflow above those components.
They usually ask for architecture diagrams, data-flow diagrams, access-control design, model inventory, risk classification, retention policy, subprocessors, patch process, incident logging, vulnerability-management process, and sample audit records.
VDF AI supports DORA-aligned programs by reducing external inference dependencies, keeping detailed ICT audit logs, supporting resilience testing, documenting model and connector dependencies, and enabling customer-controlled change windows. It does not replace your legal or supervisory compliance obligations.
VDF AI helps teams maintain AI inventories, classify AI use cases, enforce human oversight, document model and data-source choices, and produce technical evidence from governed runs. Final obligations depend on your role, use case, jurisdiction, and risk classification.
No. On-premises, VPC, and air-gapped deployments do not require third-party model APIs for production inference. If a customer chooses to connect an external model provider, that provider is customer-selected and governed by the customer’s own policy.
VDF AI is built for platform and capacity planning rather than unlimited per-token surprises. Customers run models on infrastructure they control and can combine local models, routing policy, and workload limits to make usage predictable across teams.
Trust centers should answer what runs where, which third parties touch production data, and how security-impacting changes reach regulated customers.
For on-premises, private cloud, and air-gapped deployments, inference, embeddings, private RAG, logs, and orchestration run in the customer-controlled environment.
Cloud tenancy, data center, identity provider, SIEM, secrets manager, vector database, model runtime, and storage remain under the customer’s operational and contractual control.
If a customer permits an external model, connector, or support workflow, it is explicitly configured and governed by customer policy, procurement, and risk review.
Added deeper RFP, compliance, private RAG, air-gapped deployment, and evidence-pack guidance for security reviewers.
Updated deployment diagrams and buyer questions for on-premises, VPC, sovereign cloud, and air-gapped AI agent rollouts.
Security-impacting advisories are reviewed with affected customers through the agreed support and disclosure path for their deployment model.
Send your architecture questionnaire, DORA checklist, EU AI Act evidence request, HIPAA security review, or air-gapped deployment constraints. We will answer against your environment, not a generic cloud pattern.
We will walk your architects, security, and compliance leads through the deployment model, data flow, controls, and compliance mapping for your environment — and answer your security questionnaire against a concrete reference architecture.