ENGARPAI Talk to sales

Self-hosted · Framework-agnostic · Audit-ready

Enterprise Governance,
Risk and Assurance platform for your AI systems

ENGARP AI is the system of record for every AI system you run - inventory, risk, controls, evidence, independent testing, findings, approvals and release gates in one auditable chain. Deployed entirely inside your infrastructure, with no vendor SaaS dependency for core governance.

Talk to sales See the platform
57
baseline controls across 12 categories
9
importable framework content packs
0
internet dependencies for core governance
Your AI estateLive
AI agentsML modelsLLMsAI apps
1Register2Assess3Govern

ENGARP AI · one internal control library

ControlledEffectiveEvidencedHashedGate PASSExplained
Registered48Chain breaks17Gates passed41
Deployable entirely inside client infrastructure · no vendor SaaS dependency for core governance · air-gap supported

The assurance chain

A system is only as governed as its weakest link.

Most tooling asks “are we compliant with X?” and answers with a score. ENGARP AI asks a harder question: what can this AI system actually evidence today? The answer is a chain - and the console is built around showing you exactly where it breaks, for the whole portfolio and for one system.

Assurance chain · portfolio view

Systems that have cleared each link. Counts are records that exist - nothing here is scored or weighted.

48 systems in scope

  1. 48Registered
  2. 48Risk assessed
  3. 48Controls scoped
  4. −1731Controls effective
  5. 31Evidence attached
  6. −724Independently tested
  7. 21Findings resolved
  8. 19Approved
  9. 19Release gate
Chain breaks at “Controls effective” 17 systems have not cleared it - controls are scoped but not yet evidenced as effective.AIS-0104AIS-0117AIS-0122+14 more

Nothing is scored

A framework is a mapping onto one internal control library - never a parallel programme of work, and never a percentage.

Nothing is rewritten

The audit trail is append-only. Evidence versions accumulate; a change that invalidates a gate PASS raises new work rather than editing the old verdict.

Nothing is claimed for you

The platform will never tell you that you are compliant. That is a legal conclusion someone else draws - it tells you what you can evidence.

The console

Every link in the chain has a screen behind it.

Fifty-four server-rendered routes, one dense high-signal console. Semantic colour is reserved strictly for state - a red chip means the same thing everywhere in the product. An auditor signs in to the same console read-only, and an engagement exports exactly the package they asked for.

An inventory that is a register, not a spreadsheet.

Every AI system carries an owner, a business unit, a lifecycle status, an autonomy level and a citable code. Onboarding is resumable and the Copilot can draft the record - a human still submits it.

  • Autonomy levels A0-A4 drive control scoping, not a free-text field
  • Business-unit scoping enforced on every request
  • Lifecycle from Draft through Production, Suspended and Retired
engarp.internal.acmefs.example
AI SystemsFilterRegister systemGA

Registered

48

In production

19

Under assessment

11

Tier 4 · Critical

6

AI systems

48 records · 4 shown
CodeSystemOwner · unitLifecycle
AIS-0104Claims triage assistantR. Mehta · ClaimsRemediation
AIS-0117Underwriting risk copilotJ. Okafor · UnderwritingUnder Assessment
AIS-0122Collections outreach agentL. Sandoval · CollectionsProduction
AIS-0131KYC document extractionP. Lindqvist · Financial CrimeApproved

Framework content packs

Adopt a framework without starting a second programme.

Frameworks load as versioned content and map onto one internal control library. Add a standard and you get a new view of the work you already do - not another backlog. Pick a pack to see how its requirements resolve.

Available packs

NIST AI RMF 1.0

Loaded as versioned content. Subcategories map onto your control library.

Mapped, not scored
RequirementMapped controlsCoverage
GOVERN 1.1

Legal and regulatory requirements are understood and documented

CTL-AI-001CTL-AI-003Covered
MAP 2.3

Scientific integrity and TEVV considerations are documented

CTL-AI-031TST plansEvidence pending
MEASURE 2.11

Fairness and bias are evaluated and results documented

CTL-AI-031Partially covered
MANAGE 4.1

Post-deployment monitoring plans are implemented

CTL-AI-052CTL-AI-054Covered

Import is dry-runnable, versioned, and reversible. Superseding a pack version raises change-impact work rather than silently re-mapping history. Packs are pulled on a schedule you own and verified against a trust store you pinned - the registry never holds a credential into your deployment, and anything you authored is skipped and reported, never overwritten.

A copilot that drafts. It never decides.

The Governance Copilot explains a control, drafts an assessment, summarises a finding, and cites the records it used. Accepting a risk, approving a control, accepting evidence, waiving a requirement and approving a release stay explicit human actions - enforced in code, not in policy.

  • Answers from your own records when no model is configured.
  • No external AI provider is required at any point.
  • Every draft is attributed and every acceptance is a named human act.

Separation of duties, enforced server-side.

Thirteen roles and business-unit scoping, checked on every request. Boundaries are real, not a matter of hiding buttons: an Auditor cannot edit anything, a System Owner cannot approve their own risk acceptance, and a unit-scoped analyst sees only their unit's systems.

Platform AdministratorAI Governance AdministratorGovernance AnalystAI System OwnerControl OwnerEvidence ContributorAI Security ReviewerPrivacy ReviewerRisk ReviewerBusiness OwnerApproverAuditor · read-onlyViewer

Deployment

It runs in your infrastructure. All of it.

Docker Compose for evaluation, a Helm chart for Kubernetes and OpenShift, and documented air-gapped installation. Governance, workflow, assessment, evidence and audit require no internet connection - ever.

Your infrastructure

Air-gap capable

Next.js console

54 server-rendered routes

NestJS API

The system of record

Worker

Optional · schedules & queue

PostgreSQL 16

82 governance entities

S3-compatible store

Hashed evidence objects

Your model provider

Optional · or none at all

All governance data - inventory, risk, controls, evidence, findings, approvals, audit trail - stays inside this boundary.

Licences
Content packs
Opt-in telemetry
Granted bundles

Vendor control plane

  • Offline-verifiable licences & trust keys
  • Content pack registry
  • Support desk & onboarding runs
  • Opt-in fleet telemetry

The control plane cannot read your governance data. There is no capability in that API to grant - support works through bundles you choose to share.

Suspension is read-only, never dark

You are under retention obligations for these records. A billing dispute must not be able to create a compliance breach.

Every telemetry field is registered

A published register lists every field that ever travels from your deployment to the vendor. Telemetry is opt-in.

Documented operations

Installation, sizing, backup, DR, upgrade, air-gap, observability, hardening and troubleshooting - written, not implied.

Security & assurance of the platform itself

A governance platform should be able to evidence itself.

The same standard we ask of your AI systems. Checks that stop passing are treated as control regressions, and the threat model states its residual risks plainly rather than omitting them.

24

Executable security checks

Each asserts a refusal. Safe to run against a live deployment, and they belong in your CI.

370

Unit tests on decision logic

The engines that classify risk, scope controls and evaluate gates are tested as pure logic.

AA

WCAG AA in browser journeys

Role-based end-to-end journeys assert accessibility alongside behaviour, not after it.

∞

Append-only audit trail

Hashed, versioned evidence. Nothing is silently overwritten and no verdict is edited after the fact.

Published threat model

With its residual risks stated, not buried. A threat model that only lists mitigations is marketing.

Generated evidence, not remembered claims

The entity model and end-to-end validation report are regenerated from the live system and fail the build when they drift.

An implementation-status register

Implemented, partial, or planned - with the evidence for each. Read it before relying on a capability.

Licensing

Annual licence, per deployment.

Licences are offline-verifiable and never call home to work. Pricing is quoted against deployment scale and support commitments - talk to sales and you will get a number, not a funnel.

Programme

One governance programme, one environment

For a first portfolio under formal governance. Single deployment, Docker Compose or a single-node Kubernetes namespace.

  • Full governance chain, all 16 phases
  • 57 baseline controls + 9 content packs
  • Standard support desk
Request a quote
Most deployments

Enterprise

Multi-unit portfolio with real oversight

For organizations running AI across business units with internal audit in the loop. Helm on Kubernetes or OpenShift, HA-ready.

  • Business-unit scoping and 13 roles
  • Auditor workspace & audit engagements
  • Integrations & continuous monitoring
  • Guided onboarding runs
Talk to sales

Regulated & air-gapped

No egress, no exceptions

For sovereign, classified and regulated estates. Offline licence verification, offline content pack delivery, no telemetry.

  • Air-gapped installation values for Helm
  • Telemetry disabled by contract and by build
  • Named engineering contact and DR review
Start a conversation

Stated up front

Five things this product will not do.

These are design decisions, not gaps. If a vendor promises you the opposite of any of them, ask how.

  1. 01

    It will not tell you that you are compliant.

    It tells you what is mapped, covered, partially covered, evidence pending, or not assessed. Compliance is a legal conclusion someone else draws.

  2. 02

    It will not let an AI decide anything.

    The Copilot explains and drafts. Accepting a risk, approving a control, accepting evidence, waiving a requirement and approving a release are explicit human actions, enforced in code.

  3. 03

    It will not shut your AI systems down.

    A governance hold records an organizational decision and notifies the owners. The platform holds no runtime privileges over your systems - claiming otherwise would give false assurance at the worst possible moment.

  4. 04

    It will not rewrite history.

    The audit trail is append-only, evidence versions accumulate, and a change that invalidates a previous gate PASS raises new work rather than editing the old verdict.

  5. 05

    It will not require an internet connection.

    Not for governance, workflow, assessment, evidence, or audit. Air-gapped installation is documented, supported, and tested.

Bring the chain to your portfolio.

A technical briefing runs ninety minutes: the governance chain against one of your real AI systems, the deployment topology against your estate, and the threat model with its residual risks in the open.

01Ninety-minute briefing with an engineer, not a demo reel.

02Self-hosted evaluation in your environment, seeded with a realistic in-flight portfolio.

03A quote against your deployment scale - offline-verifiable licence, annual term.

Talk to sales

We will not add you to a drip sequence. One reply from a person, then it is your call.