Software that works
where the internet does not.
Most commercial software is built on assumptions a sensitive environment cannot grant. The Praesidium Suite is Salian Defense's software line for the other case: classified, CUI, disconnected and accredited space, where inference runs on hardware you own, egress is a decision rather than a default, and every sensitive action stops at a person.
What is the Praesidium Suite?
The Praesidium Suite is Salian Defense's line of software built for environments that cannot reach the internet: secure facility scoping and cost modeling, ICD-705 design validation, controlled AI workflow orchestration, AI boundary policy, retrieval, and partner-delivered inline encryption for legacy equipment.
Every product in it shares the same architecture: local model routing, egress control, signed bundles across an air gap, human approval gates and an audit trail written as the work happens. Products are described here by what they do and who uses them. Which of them belong in a given environment is a conversation, not a catalog.
The assumptions ordinary software will not name.
Nothing on this list is unreasonable. Each one is a sensible default for a company with an office, a browser and a corporate network. Each one is also a hard dependency that a sensitive, controlled, disconnected or accredited environment will refuse, and most of them are invisible until the software is already bought.
- Constant internetA live connection at run time for licensing, activation, feature flags, content and configuration. The enclave has none, and a "degraded mode" written for a bad hotel connection is not the same thing as a permanent one.
- SaaS identityAuthentication delegated to an external identity provider. Inside the boundary the directory is local, the accounts are locally administered, and no request may leave to prove who a user is.
- External model APIsInference sent to a vendor's endpoint. That is a disclosure event, and no wording in a terms-of-service page converts it back into a local computation.
- Cloud telemetryCrash reports, usage analytics and performance traces sent outward by default. Even content-free telemetry is unreviewed egress, and unreviewed egress is the thing the boundary exists to prevent.
- External loggingLogs shipped to a hosted aggregator. The audit trail for classified work has to live where the work lives and be producible to the people who ask for it.
- Vendor-controlled updatesA background updater that changes the software on the vendor's schedule. In an accredited system a change is an event the security officer authorizes, records and can reverse.
- Open network accessFree outbound reach to registries, CDNs and third-party services at run time. Dependencies that resolve over the network at start-up simply do not resolve here.
The result is familiar. A capable product is procured, and then discovered to be unusable inside the boundary it was bought for. The remedy is not a harder deployment. It is software whose architecture never assumed any of the above.
Five decisions every product in the suite makes the same way.
These are architectural properties of the software, stated as design, not as claims about anyone's deployment. They are what makes the products behave predictably when the network is not there.
Every product resolves a data profile at start-up and routes inference accordingly. Under a disconnected profile the model is served inside the boundary and an outbound call is refused rather than retried. An external provider is selectable only under an unclassified profile and only behind an explicit gate an owner has to set.
Outbound calls pass a guard at call time rather than relying on the network to block them. The failure mode is a refusal that is visible in the logs, not a silent request that a firewall happens to drop. What is allowed to leave is a configured list, not whatever a dependency decides it needs.
Software, dependencies, model weights and migrations cross the air gap as one signed bundle, verified on the inside before anything is applied. Nothing is pulled from a registry at run time, so an update is a scheduled, recorded, reversible event rather than something that arrives on a vendor's schedule.
Sensitive actions decompose into steps, and a step that matters waits on a person to approve it before it proceeds. Automation is used to prepare and sequence work, never to take a consequential action unattended.
State transitions, approvals and policy verdicts are recorded as they occur, with classification carried alongside them, so the record is a by-product of the work rather than a reconstruction assembled later for an assessor.
Hosted and unclassified, customer-hosted so the software stays outside your CUI boundary, or fully disconnected inside accredited space. The profile is resolved at deploy time and changes what the software will and will not attempt.
The system view.
A cleared program moves through the same arc whether or not anyone draws it: find the work, qualify it, scope the environment, design and validate it, build it and carry it to accreditation, then operate inside it. The map below places each piece of Salian's work at the point in that arc where it belongs.
- Find the work RFI Hawk Salian-operated venture, hosted
- Qualify it Capture and advisory Direct Salian service
- Scope the environment Vates Cognition Praesidium Suite
- Design and validate Groma Praesidium Suite
- Build and carry to accreditation Salian facility delivery Direct Salian capability
- Operate constrained workflows Legatus Praesidium Suite
- Define AI boundaries Vigil Praesidium Suite
- Retrieve knowledge Hawk RAG Praesidium Suite
- Protect legacy links Partner hardware Delivered through a partner
- Monitor compliance and program state Project Hawk Salian-operated venture, customer-hosted
Read this as a relationship map, not a deployment diagram. It shows where each piece sits relative to the others and where the handoffs are. It is not a claim that every stage runs in one installation, and no environment needs all of it. Most need two or three of these and nothing else.
What each one does, and who uses it.
Five software products and partner hardware. Each is described below by its function, its user, the environment it is built for, and how it is deployed.
Vates Cognition
Turns a short program intake into a budgetary rough order of magnitude for secure-facility work, with a compliance engine behind the number so the range moves as the requirement is described rather than as a spreadsheet is nudged. The estimate tightens as the intake is completed.
- Who uses it
- Program offices, facility and security officers, and capture teams sizing a budget before design money is committed.
- Built for
- Organizations that need a defensible facility number early, without sending program parameters to a hosted estimator.
- Deployment
- On-premise, self-hosted. Any outbound AI call sits behind a kill switch and a separate data-retention gate that an owner has to set explicitly; nothing leaves the process until both are on.
- Try it
- The free Project Estimator on this site runs the same engine.
Groma
Reconstructs an ICD-705 model from a floor plan, a point-cloud scan or a written description, then checks it against buildingSMART IDS 1.0 plus nine deterministic checks and reports what fails and why. Every output is a draft for a credentialed human to review and ratify.
- Who uses it
- Facility security officers, security specialists and architects running a pre-inspection self-check or triaging a facility they inherited.
- Built for
- Work on real facility drawings, which are exactly the documents nobody should upload to a commercial service.
- Deployment
- On-premise, with an air-gapped profile that routes inference to a model served inside the boundary. Under that profile an egress guard refuses external calls at call time.
- Related
- SCIF accreditation covers the process the checks are written against.
Legatus
Decomposes an analyst's stated intent into a multi-step workflow, runs each step through a controlled adapter rather than a vendor SDK, and stops at a human for a signed approval before any sensitive action proceeds. Classification rides every state transition, and the graph fails closed rather than leaking at run time.
- Who uses it
- Mission, analytic and engineering teams that need multi-step automation inside a boundary where an unattended action is not acceptable.
- Built for
- Disconnected operation inside accredited space. Built to the DoD Impact Level IL4 to IL6 control set, which is a design target rather than a determination.
- Deployment
- On-premise Kubernetes or Compose, crossing an air gap as a single signed bundle. Model backends are pluggable, with a locally served model as the default posture.
Vigil
Defines and evaluates the boundary policy an AI system must respect: what to allow, redact or block. A deterministic detector set covers identifiers, payment numbers and keys, a policy engine runs in advisory or enforcing mode, and a redaction transformer applies the verdict. Verdicts are content-free by construction, so the record of a decision never becomes a second copy of the sensitive text.
- Who uses it
- Engineering teams and security officers who have to state, in writing and in code, what an AI system is permitted to see and emit.
- Built for
- Environments where the AI question is an architecture question: which boundary the model sits behind, and what is allowed to cross it.
- Deployment
- An embedded library inside the host application rather than a separate service, so it adds no network hop and no external dependency of its own.
Hawk RAG
The retrieval layer shared by the Hawk products. It fuses vector, knowledge-graph and lexical retrieval by reciprocal rank, so a query returns connected facts rather than a ranked list of documents that each mention the search term.
- Who uses it
- Analysts working across large procurement, regulatory and program corpora, through the products that embed it rather than as a tool of its own.
- Built for
- Retrieval that has to run wherever the corpus is allowed to live, including inside a boundary the corpus may not leave.
- Deployment
- A package that runs inside the host product's boundary on Postgres. It runs end to end without any paid model call unless that is explicitly enabled.
Pythian
FPGA-based inline encryption hardware for operational technology and industrial control networks. It installs between devices, with no code changes, no operating system modification and nothing on the endpoint, for legacy and vendor-locked equipment that will never accept software.
- Who uses it
- Facility and OT engineers protecting links between equipment that cannot be patched, replaced or instrumented.
- Built for
- Links that carry sensitive traffic between devices whose vendors ended support long ago.
- Deployment
- An inline appliance, delivered through a partner. Salian is the integrator rather than the manufacturer and is not presently under license. Validation status, form factor and supply-chain provenance are provided under NDA against a specific requirement.
The ventures are separate.
Three products carry the Hawk name and are operated by Salian Defense as ventures rather than sold as part of the Praesidium Suite. They appear on the relationship map because they occupy real positions in the same lifecycle. Each can be bought on its own, without buying anything else.
Federal opportunity discovery and proposal drafting, scored against a company's capability profile. Runs hosted at rfihawk.io with a free tier. Used by capture managers, proposal writers and compliance officers.
Continuous compliance automation for CMMC, NIST SP 800-171 and NIST SP 800-53 posture, deployed on customer infrastructure so it stays outside the CUI boundary rather than inside it.
Public-data intelligence on United States public water systems, scoring chemical-supplier exposure and discharge posture. Runs hosted at waterhawk.io for suppliers, utilities and infrastructure investors.
Do you need custom software?
This is the question most technology conversations with Salian actually start from, and the honest answer is usually not to buy anything. Custom development is the most expensive thing on this page to own, and most environments land on one of the four cheaper outcomes above it. Work down the list. Stop at the first one that is true.
-
Configure the tools you already have
The capability is already licensed and switched off, misconfigured, or deployed in a profile that does not match the environment. This is the most common finding and the cheapest fix, and it very often ends the conversation.
You are here if the gap is a workflow nobody has set up, not a function nobody has written. -
Integrate the systems you already run
Each system does its part and none of them speak to the next one, so people carry data between them by hand and the errors live in that gap. The work is an interface and a schedule, not a product.
You are here if the answer to "where does that number come from" is a person and a spreadsheet. -
Build a controlled workflow layer over them
The underlying systems are fine, but the sequence of steps between them has to be governed: classification carried through, sensitive actions gated on a human approval, every transition recorded. That is a thin layer over what exists rather than a replacement for it.
You are here if the requirement is about who approves what and what gets recorded, not about new functionality. -
Evaluate on-premise software
A product already solves this, and the only real question is whether it survives the environment: no external identity, no external model call, no telemetry, no run-time network reach, updates as signed bundles. The Praesidium Suite exists for the cases where the answer is yes.
You are here if the function is common but every candidate you have priced assumes the cloud. -
Custom development may be justified
The workflow, the classification boundary or the disconnected posture genuinely has no product behind it, and the cost of owning code for years is smaller than the cost of continuing without it. Salian builds in this case and says so plainly when a case has not been made.
You are here if you have worked through the four outcomes above and the requirement is still standing.
Salian will tell you which of the five you are in, including when the answer costs nothing. The engagement that starts with an honest "you do not need this" is the one that makes the next one worth having.
Air-gapped software, answered.
What does air-gapped software actually mean?
Do these products call an external model API?
Is the Praesidium Suite accredited?
How do updates cross an air gap?
Are RFI Hawk, Water Hawk and Project Hawk part of the Praesidium Suite?
We think we need custom software. Do we?
Talk through the environment.
Describe the boundary, the workflow and the constraint, and we will tell you which of the five outcomes you are in and which pieces, if any, apply. The first conversation is unclassified, and no classified detail is required to begin. If the answer is that you should buy nothing, that is the answer you will get.