Sovereign AI
AI that runs inside your environment and inside your workflow.
Nexus uses self-hosted models for classification, extraction, summarization and bounded reasoning, with deterministic rules and human approval deciding what actually happens to your systems.
AI as one governed stage
The model interprets. The workflow decides what is permitted, and a person confirms anything consequential before a system is changed.
- model execution
- Self-hostedmodel execution
- required for inference
- No egressrequired for inference
- capable deployment
- Air-gappedcapable deployment
- prompt and output history
- Full runprompt and output history
Sovereignty
Where the model runs is an architectural decision, not a setting.
For regulated, defense-adjacent and public sector work, the question is rarely which model is best in a benchmark. It is whether the data can leave the building at all.
Models run where you run
Inference executes on infrastructure you control, whether that is a data center rack, a private cloud tenancy or a sovereign region under a specific jurisdiction.
No external LLM API dependency
Workflows do not require a call to a third-party model provider. There is no vendor endpoint in the execution path that could change terms, pricing or availability.
No data egress required
Documents, prompts, retrieved context and model outputs stay inside your network boundary. Nothing has to leave the environment for the workflow to complete.
Air-gapped operation
The same workflows run in disconnected networks. Models and knowledge bases are loaded into the environment rather than reached across the internet.
Components
What you operate.
Sovereign AI is not a single toggle. It is a small set of components your platform team runs and your workflow authors consume.
Model registry
Register the models available to your workflows and control which ones a workspace may use.
Inference endpoints
Configure the endpoints that serve your models so workflow nodes call a governed target rather than an ad hoc address.
Knowledge bases
Curate private document collections that a workflow can retrieve from at run time.
Retrieval in workflows
A node retrieves relevant passages and passes them to the model as context, keeping answers grounded in your own material.
Prompt configuration
Prompts are part of the workflow definition, versioned with the revision rather than pasted into a chat window.
Structured output
Model responses are shaped into typed fields the rest of the workflow can branch on and write back.
Guardrail nodes
Validation, thresholds and policy checks run around the model call rather than being trusted to the prompt.
Workspace scoping
Model access, knowledge bases and credentials follow the same workspace boundaries as the rest of the platform.
Separation of concerns
Deterministic rules and model judgment do different jobs.
Most disappointing AI projects blur these two. Nexus keeps them explicit, because they have different failure modes and deserve different controls.
Deterministic rules
Anything that must be correct every time
- Entitlement, eligibility and policy checks
- Arithmetic, totals, tolerances and thresholds
- Required fields, formats and referential integrity
- Which system is called and with which credential
- Routing, retries and exception branches
Model judgment
Anything that requires reading meaning
- Interpreting unstructured text and documents
- Classifying an ambiguous case
- Summarizing history for a reviewer
- Proposing a likely resolution
- Drafting language a person will confirm
System actions remain governed by workflow logic in both cases. A model result changes which branch is taken. It does not change what a branch is allowed to do.
Applied AI
Where AI earns its place in an operation.
These are the tasks where a model reliably reduces manual reading and typing, provided the surrounding workflow validates and records the result.
Classification
Assign a case, document or alert to a category so downstream deterministic branches can route it.
Extraction
Pull structured fields out of unstructured input such as invoices, contracts, tickets and email threads.
Summarization
Condense long correspondence, log output or case history into the context a reviewer needs to decide.
Recommendation
Propose a next action or a likely resolution, which the workflow treats as a suggestion and not an instruction.
Agent-like reasoning inside a workflow
Break a bounded task into steps within a defined set of permitted actions, with the workflow retaining control of what can actually execute.
Drafting
Produce a first version of a response, a summary note or a case record for a person to review and correct.
Worked example
AI inside a governed run.
A document arrives, deterministic checks run first, the model interprets what rules cannot, and a person confirms before anything is written back.
- 01
Trigger
A supplier email with an attached document arrives and starts the workflow.
- 02
Integration
Connector actions fetch the related purchase order and vendor master data from the systems of record.
- 03
Rules + Code
Deterministic checks confirm the vendor exists, the totals reconcile and the document format is valid. Cases that fail hard rules never reach the model.
- 04
Sovereign AI
A self-hosted model extracts line items and summarizes the free-text notes, grounded by retrieval from your internal policy knowledge base.
- 05
Human Approval
A reviewer sees the extracted fields, the source document and the model output, then confirms, corrects or rejects.
- 06
System Action
The confirmed result, not the raw model output, is written to the operational system.
- 07
Audit
Model, prompt, retrieved context, response, reviewer identity and final decision are retained with the run.
- 08
Exception path
Low-confidence or policy-flagged cases route to escalation instead of proceeding to a write-back.
Guardrails
Controls around the model, not inside the prompt.
A prompt is guidance. A guardrail is enforcement. Nexus places the enforcement in the workflow, where it can be reviewed and audited.
The model does not hold the credentials
System actions are executed by workflow nodes with vault-backed credentials. A model output can influence which branch runs, but it does not itself gain access to a target system.
Deterministic checks bound the outcome
Thresholds, allowed values, required fields and business rules are evaluated in code and configuration around the model call, so an unexpected response fails validation rather than propagating.
Approval on consequential steps
Where an action has financial, contractual or operational weight, a person confirms before it executes. The workflow enforces that, not a convention.
Scoped knowledge
A workflow retrieves from the knowledge bases it is permitted to use, which keeps a model from grounding an answer in material the requesting team should not see.
Confidence and fallback routing
Uncertain results can be routed to enrichment, a second check or a reviewer instead of being treated as a settled answer.
Traceability
Every AI step leaves a record you can inspect.
If a model influenced an operational outcome, you should be able to reconstruct exactly what it saw and what it returned.
Prompt and input
The prompt used, the inputs supplied by earlier nodes and the retrieved context are retained with the execution rather than reconstructed from memory.
Model and endpoint
The run records which registered model and endpoint served the request, which matters when a model is updated and behavior changes.
Output and downstream use
The response is stored alongside the branch it selected and the values it contributed to the eventual system action.
Human decision
Where a reviewer approved, corrected or rejected the result, the decision, the person and the reason are part of the same record.
Private knowledge use
Retrieval from internal knowledge bases is logged with the run, so it is clear which internal material informed an answer.
Operation
Operating responsibilities.
Deterministic checks, scoped credentials and human approval govern how model outputs enter a workflow.
Checks around model output
Model outputs pass through deterministic checks and human approval according to the workload and its consequences.
Deployment-specific assurance
Your compliance position depends on deployment, configured controls and audit evidence. Nexus records the workflow decisions and actions those reviews need.
Credentials and approval stay governed
Models do not hold credentials or approve their own actions. Consequential steps remain governed by workflow logic and, where appropriate, a named reviewer.
Architecture session
Put AI where the data already lives.
If your material cannot leave the environment, the model has to come to it. Bring us the process and we will show how self-hosted inference, private retrieval and human approval fit into one governed run.