AIMAILABS platform architecture

Compile the rulebook.
Execute the decision.

The platform separates probabilistic AI perception from deterministic operational execution. Models can change. The governed decision process remains stable, testable and traceable.

Assess a Workflow →See Decision Patterns

Architecture at a glance

Two engines. One controlled evidence-to-decision pathway.

Compiler

Operational Knowledge Compiler

Converts governed human knowledge into executable decision packages before the workflow goes live.

Source registry

Preserves document identity, revision, ownership, effective date and source trace.

Ontology mapping

Normalises entities, qualifiers, aliases, allowed values and relationships.

Decision intermediate representation

Structures inclusion, exclusion, calculation and recommendation logic.

Evidence policy compiler

Defines what is visually observable, document-derived, sensor-derived or measurement-required.

Governance compiler

Encodes approval authority, escalation, provisional status and human-only decisions.

Static tests and release

Finds contradictions, duplicate rules, missing evidence and version incompatibility.

Runtime

Governed Decision Runtime

Applies an approved package to a verified evidence state during operation.

Evidence intake

Receives images, documents, sensors, forms and user observations with provenance.

Field-level verification

Accepts, corrects, rejects, marks unresolved or requests external confirmation.

Canonical normalisation

Only ontology-valid fields may cross the structured decision boundary.

Constraint resolution

Eliminates incompatible decision records before any optional ranking.

Assessment and action

Calculates governed ratings, priorities, recommendations and output states.

Provenance and approval

Links the result to evidence, rule version, rejected alternatives and authorised reviewers.

The LLM is replaceable. The compiler, evidence schema, rules engine, governance and traceability remain AIMAILABS platform assets.

What gets compiled

From human guidance to a deployable decision package.

Source knowledge

Human-readable inputs

Procedures and SOPs
Technical standards
Decision matrices and checklists
Historical approved cases
Expert judgement and approval rules

Compiler output

Machine-executable package

Canonical entity and qualifier registry
Required and prohibited constraints
Evidence determinability policies
Calculations and action codes
Version, approval and source trace

Deployment control

Tested release

Positive and negative test cases
Contradiction and conflict checks
Technical authority approval
Digital release status
Rollback and revision management

Evidence determinability

The runtime controls what AI is allowed to know.

A machine-generated value cannot influence a decision merely because a model produced it. Each field carries a determinability policy and verification state.

Visual direct

A condition is directly observable in the evidence and may be verified visually.

Context dependent

The observation requires scene context, component identity or multiple evidence views.

External confirmation

The field requires a measurement, sensor, document or authorised human input.

Not machine determinable

The runtime rejects model-estimated values and blocks unsupported finalisation.

Example: unsupported inference

AI proposes a remaining thickness value from a photograph.

The rule package classifies thickness as measurement-required.

→

Runtime response

Discard the unsupported value.

Keep the assessment provisional and request UT confirmation.

Decision governance

Not every result is allowed to become a final decision.

Runtime state	When it applies	Permitted outcome
Auto-select	All required evidence is verified and the rule explicitly permits automation.	System may select the record while preserving an audit trail.
Provisional	A likely rule is identified but an external measurement or confirmation remains outstanding.	Draft result only; no final operational authority.
Candidate only	Multiple valid records remain or evidence is incomplete.	Authorised reviewer chooses among traceable candidates.
Human approval required	The rule or risk level mandates an authorised person.	System prepares the basis; human approves or overrides with justification.
Blocked	Critical evidence fails quality, ontology or contradiction checks.	Workflow stops and requests additional evidence.

Model-agnostic by design

Adopt better AI models without rebuilding the decision system.

Replaceable perception

Best-fit AI models

Cloud, local or future models can be selected by use case, privacy, cost and deployment constraints.

Stable interface

Typed evidence schema

Every model must output the same governed evidence structure before reaching the runtime.

Reproducible execution

Consistent decision result

The same verified evidence and rule version should produce the same governed decision state.

Workflow fit assessment

Can your rulebook be compiled?

The best starting point is one high-value operational workflow with identifiable evidence, decision rules and approval authority.

Assess the Workflow →

Compile the rulebook.Execute the decision.