Urquhart
Digital Services
Discuss a project
Menu
System / SARIEL UDSMode / ResearchAuthority / Human-governedAudit / Architecture principle

Urquhart Digital Services / Research

SARIEL UDS

The Intelligence Behind the System.

A modular cognitive architecture developed by Urquhart Digital Services for private, persistent and governed artificial intelligence systems.

Not another chatbot. An investigation into how intelligence can participate safely inside real operational systems.

Explore the architecture Discuss private AI systems

Why SARIEL exists

Modern organisations do not lack software. They suffer from fragmentation.

Information lives across platforms, machines and people. Traditional automation can move data. A language model can answer a question. Neither alone provides the complete architecture needed to retain context, evaluate authority, control action and preserve evidence.

SARIEL investigates that missing layer: intelligence that can exist responsibly inside a living system.

Not one model

Structured disagreement before controlled action.

SARIEL is being developed as an orchestration architecture, not one all-powerful model. A router can direct a problem toward specialised analysis, research, engineering, commercial review, criticism or verification.

These are research architecture roles. They are not presented as uniformly production-mature components.

Architecture model / Research

Reasoning is separated from authority.

Council coordinates specialised intelligence. Memory, Critic and Context inform reasoning. Sentinel and Gate control the authority boundary before Action. Ledger preserves evidence afterwards.

SARIEL Engine / Research modules

Six responsibilities surrounding intelligence.

Each module describes a deliberate architectural responsibility. Their significance lies in separation: memory is not authority, confidence is not permission, and action is not complete without evidence.

01Architecture component

Memory

Preserve context beyond a single conversation.

  • Operational state
  • Historical events
  • Verified knowledge
  • Assumptions
  • Rules
  • Unresolved uncertainty

Memory is treated as a governed system, not an uncontrolled archive.

02Architecture component

Council

Route work between specialised models, reasoning processes and tools.

  • Analysis
  • Planning
  • Technical design
  • Research
  • Commercial review
  • Verification

Different problems require different forms of intelligence.

03Architecture component

Critic

Challenge reasoning before confidence becomes action.

  • Assumptions
  • Logical gaps
  • Conflicting evidence
  • Incomplete information
  • Unsafe actions
  • Excessive confidence

A convincing answer must not be mistaken for a correct one.

04Architecture component

Sentinel

Control the boundary between reasoning and action.

  • Permissions
  • Production data
  • Financial consequences
  • Private information
  • Reversibility
  • Human approval

SARIEL may recommend an action. Authority determines whether it may act.

05Architecture component

Gate

Create deliberate control points before consequential action.

  • Additional evidence
  • Second evaluation
  • Human confirmation
  • Rollback
  • Authority level
  • Audit requirements

What the system believes is not the same as what the system is allowed to do.

06Architecture component

Ledger

Preserve evidence around meaningful decisions and actions.

  • Information used
  • Conclusions
  • Responsible module
  • Remaining uncertainty
  • Authorisation
  • Validation

Accountable automation.

Private by design

Control the boundary, not just the model location.

Different workloads require different architectures. SARIEL can investigate local inference, selected external models, private infrastructure, controlled data boundaries, hybrid systems and local knowledge stores.

Local operation does not automatically create security. It creates greater control over where information is processed. Access, logging, updates, backups and operating discipline still matter.

01

Local

Selected processing and knowledge remain on controlled local infrastructure where the workload justifies it.

02

External

Appropriate external models can be selected deliberately rather than assumed as the only architecture.

03

Hybrid

Data classes, capability and authority can be separated across local and external components.

Intelligence with boundaries

Capability contained by architecture.

The architecture should determine what AI can observe, remember, access, recommend, execute and when it must stop.

Built for real systems

Application areas under investigation.

These areas describe research and commercial application potential. They are not claims of completed customer deployment.

01 / Research potential

Business operations

  • Document intake
  • Workflow coordination
  • Reporting
  • Knowledge retrieval
  • Project administration
02 / Research potential

Technical systems

  • Equipment monitoring
  • Telemetry
  • Environmental sensing
  • Diagnostics
  • Controlled automation
03 / Research potential

Computing

  • Workstation health
  • Workload analysis
  • Local AI orchestration
  • Infrastructure monitoring
  • Development systems
04 / Research potential

Astronomy

  • Observatory state
  • Weather interpretation
  • Equipment coordination
  • Safety logic
  • Remote recovery
05 / Research potential

Research & development

  • Model orchestration
  • Persistent memory
  • Controlled tool use
  • Evaluation systems
  • Human-machine collaboration
Owner-supplied photograph of the UDS high-performance development workstation with internal components and red illuminated cooling fans visible
Owner-supplied development environment photograph. Hardware specifications and benchmark claims are intentionally omitted.

The physical layer

Intelligence still requires machinery.

Behind every model is physical compute. UDS develops and tests local AI, automation and technical systems using high-performance computing infrastructure designed around real workloads.

The workstation is not decoration. It is part of the engineering stack: where software, local AI and engineering meet physical compute.

Explore specialist computing AI workstation architecture

What SARIEL is today

An active internal research and development platform.

SARIEL is not presented as AGI, conscious or infallible. Its significance lies in the architecture surrounding intelligence rather than exaggerated claims about the model itself.

Active

Internal UDS research platform and development environment.

Experimental

Memory, orchestration, evaluation, controlled tool use and observability approaches.

Planned

Only separately scoped commercial applications supported by appropriate evidence and controls.

Conceptual

Application areas remain potential unless an implementation is specifically verified.

  • Persistent contextual memory
  • Multi-model orchestration
  • Specialist intelligence routing
  • Private and local AI
  • Evidence-based decision support
  • Controlled tool execution
  • Human approval gates
  • Operational monitoring and observability
  • Auditable automation

SARIEL + UDS

The research connects disciplines already present inside UDS.

It provides a foundation for systems that may need to understand information, retain context, monitor conditions, coordinate specialists, recommend actions, respect authority and explain what happened afterwards.

The principle

Most AI is designed to produce an answer.

SARIEL is being designed to participate in a system.

MemoryContextRestraintEvidenceAuthorityAccountability

SARIEL UDS

A governed intelligence architecture for systems that cannot afford blind automation.

Developed by Urquhart Digital Services.

Private AI · Persistent Memory · Specialist Reasoning · Controlled Action · Auditable Systems

Discuss a private AI or automation project