EvolvAIEvolvAI
Platform Architecture

The architecture behind intelligent enterprise AI

A hexagonal, model-agnostic orchestration engine that deploys into your cloud, assembles specialist AI teams for every department, and keeps your data exactly where it belongs — with you.

Request a DemoExplore Security

Hexagonal Architecture

Built on the ports-and-adapters pattern, the EvolvAI core engine is surrounded by interchangeable adapters for LLMs, data sources, and delivery channels. Business logic stays clean. Integrations stay flexible.

Hexagonal architecture — a Domain Core surrounded by six abstract PortsDomain CoreLLMPortConnectorPortStoragePortBillingPortAuthPortTelemetryPort
Swap Models Without Rework

Adapters decouple your workflows from any single LLM. Move from GPT-4 to Claude to Gemini without touching business logic.

Connect Any Data Source

Port interfaces let you plug in databases, APIs, SaaS tools, and internal systems through a standardized contract.

Zero Vendor Lock-in

The core engine owns no opinions about infrastructure. You choose the models, the cloud, and the integrations.

Two Ways to Work With AI

Business teams get a web dashboard. Engineering teams get native IDE tooling. Same platform, same governance, same agents — two interfaces purpose-built for how each team actually operates.

Business Teams
The Hub
The customer-facing web application — directors browse the catalog, configure agencies, monitor dashboards, and approve gates.
  • Workflow monitoring with phase-by-phase progress and outputs
  • One-click approval gates for sensitive operations
  • Real-time agent activity feeds and audit logs
  • Role-based access with SSO integration
  • Department-scoped dashboards and reporting
Engineering Teams
IDE Integration
VS Code and Cursor integration via the Model Context Protocol (MCP) — AI agents embedded directly in the developer workflow.
  • Invoke agents from your editor without context switching
  • MCP-native: works with any MCP-compatible tool
  • Code generation, review, and refactoring workflows
  • Access the full certified skill library inline
  • Same governance and audit trail as the web dashboard

Deploy to Your Cloud

EvolvAI runs inside your own cloud infrastructure. Your data never crosses a boundary you don't control. We provide the intelligence layer — you own everything else.

01

Provision

Terraform modules stand up the platform in your GCP, AWS, or Azure project. Single-tenant, your VPC, your KMS keys.

02

Configure

Compose agencies for each department from the certified library. Select models, connect systems of record, and set governance defaults.

03

Operate

Directors run governed workflows from the Hub or IDE. The platform follows the EVOLV lifecycle — pilot, scale, learn, prove value — with runbooks at every phase.

Data SovereigntyAll data stays within your infrastructure perimeter
Encryption at Rest & TransitTLS 1.3 in transit, AES-256 at rest — your KMS keys
Infrastructure-as-CodeTerraform modules for repeatable, auditable deployments
Compliance ReadyBuilt for POPIA, GDPR, SOC 2, and ISO 27001 alignment

The AI Agency Model

Every department gets a dedicated AI Agency — a curated team of specialist agents, certified skills, workflow patterns, data connectors, and governance rules. Think of it as hiring a purpose-built AI team for each function in your organisation.

What Makes Up an Agency

Agents

Specialist roles drawn from the closed roster of six

Certified Skills

Versioned methodologies with explicit acceptance criteria

Patterns

Workflow shapes and the agentic relationships within them

Channels

Where the director engages the agency — web, email, Teams

Connectors

Adapters for your systems of record — ATS, ERP, document stores

Governance Defaults

Approval gates, HITL policies, and audit boundaries

Six Specialists, One Coordinated Team

A closed roster of six named archetypes — every AI Agency is composed from these six. New capability lands as a new skill or a new pattern, never as a new agent. Arthur coordinates them all.

ArthurOrchestrator

Convenes the team for the work at hand, sequences their actions, supervises progress, and presents one unified outcome to the human director. Never produces deliverables himself.

MarcoResearcher

Discovers across internal records, document stores, knowledge bases, and the open web. Returns structured findings with citations. Gathers — never interprets.

SophieAnalyst

Reasons over evidence to produce structured judgment — comparisons, scores, anomalies, recommendations — with the evidence trail behind every conclusion.

AdaCreator

Renders analysis into the artefacts the work needs — documents, presentations, drafts, contracts, summaries — fit for the audience that will read them.

VeraReviewer

Holds every deliverable against the rules that govern it — policies, standards, compliance frameworks. Returns a verdict with the rule citation behind every finding.

FelixOperator

The only agent with side effects. Sends, books, posts, files, calls APIs — every action verifiable, logged, idempotent where possible, recoverable when not.

Skills are structured methodologies with explicit acceptance criteria — organised across seven categories: cognitive, communication, compliance, finance, integration, production, and talent. Every skill is versioned, reviewed, and passed through a certification pipeline before it reaches your platform.

How a Request Flows

Three tiers. One predictable shape.

Arthur classifies every incoming request before any work begins. The tier determines the orchestration shape: direct answer, single-agent delegation, or planned multi-phase execution with explicit human approval.

Tier 1

Direct

Answerable from context Arthur already has — no specialist invocation needed.

"What's the status of the candidate I asked about yesterday?"

Shape

Path: Arthur (Orchestrator).
Tier 2

Single-agent

One specialist completes the work end-to-end. Single-hop delegation via the routing pattern.

"Score this CV against the role spec." → Arthur routes to Sophie.

Shape

Path: Arthur (Orchestrator) → Sophie (Analyst).
Tier 3

Multi-phase

Multiple specialists in sequence with an upfront plan. Arthur produces phases, dependencies, milestones, and risks; gets explicit approval; then delegates phase by phase.

"Run the quarterly board pack." → Arthur plans → Marco gathers → Sophie analyses → Ada drafts → Vera reviews → Felix files.

Shape

Path: Arthur (Orchestrator) → Marco (Researcher) → Sophie (Analyst) → Ada (Creator) → Vera (Reviewer) → Felix (Operator).

Tier 3 work pauses for explicit director approval before any phase begins. Felix, the only side-effecting agent, also requires per-action approval by default — every email, file, or API call is governed at the gate.

Patterns

Two vocabularies, one composable system.

A template uses one workflow pattern as the spine and one or more agentic patterns to express the relationships within it. Both vocabularies are enum-enforced in the agency schemas.

Agentic

How agents relate at a step or step-cluster.

chaining

Sequential agent calls where each agent’s output feeds the next. The simplest pattern. Use when the work is genuinely sequential and each handoff is well-defined.

routing

Arthur classifies the request and dispatches it to exactly one specialist. Single-hop delegation. Default for Tier 2 requests.

orchestrator-worker

Arthur decomposes work into typed subtasks, dispatches each to the appropriate specialist, and synthesises the returned outputs. Default for Tier 3 multi-phase workflows.

evaluator-optimizer

A producer generates output; an evaluator scores it against criteria; the producer revises; loop until cleared. Most commonly Ada↔Vera revision loops.

parallelization

Independent subtasks run simultaneously across multiple agent invocations; results are aggregated. Use when subtasks are genuinely independent and calendar time matters.

Workflow

What kind of work the agency does end-to-end.

create

Produce something new — a deliverable, a plan, a design — from a brief and source material. Default for content-production agencies.

assess

Evaluate a defined subject (candidate, supplier, document, option) against a rubric and return a structured verdict with rationale and dissent.

resolve

Handle an open issue — incident, complaint, dispute, exception — from triage through diagnosis, action, and close-out.

optimize

Improve a working but suboptimal artefact, process, or output. Distinct from resolve (issue handling) and create (new production).

operate

Run a recurring operational task with consistent execution and audit per cycle. Cycle telemetry feeds the trend signal that powers optimize.

discover

Open-ended search and pattern-finding when the question itself is being formed. The exploratory shape; produces direction more than deliverables.

Ready to see it in action?

Discover how EvolvAI deploys securely into your environment, or explore pricing for your organisation.

Explore SecurityView Pricing