Skip to main content

Platform Architecture

Velatir consists of five main components that work together to provide comprehensive AI governance:

How Components Work Together

Trace Created

Your AI system sends a trace through API, SDK, or integration. Traces can be Inlet (incoming request), Response (outgoing), or Signal (passive event).

Session Grouping

Traces with the same session ID are grouped together, providing full context for an interaction.

Policy Assessment

Each trace is evaluated against active policies (pre-built and custom rules).

Workflow Execution

If a workflow is configured, triggers are evaluated. Matching triggers execute the workflow nodes.

Human Intervention

If the workflow hits a human intervention node, a review task is created and sent to the configured channel.

Decision & Completion

Human approves, declines, or requests changes. The trace is marked complete and logged.

Key Principles

1. Traces Are the Foundation

Every AI interaction is captured as a trace. This gives you complete visibility into what your AI systems are doing.

2. Sessions Provide Context

A single user conversation or workflow run generates multiple traces. Sessions group these together so reviewers see the full picture.

3. Workflows Control the Flow

Visual, node-based workflows let you define exactly how traces should be handled based on policy results, risk levels, or custom conditions.

4. Review Tasks Are Created When Needed

Not every trace needs human review. Review tasks are only created when a workflow’s human intervention node is triggered.

5. Work Where You Work

Approvals happen in your existing tools - Slack, Teams, email, or the web dashboard.

Trace Lifecycle

Every trace follows this pattern:
  1. Received - Trace arrives from your system
  2. Assessed - Policies evaluate the trace
  3. Workflow Executed - Triggers checked, nodes executed
  4. Completed/Rejected - Final state based on workflow outcome or human decision

Trace Directions

DirectionDescriptionExample
InletIncoming request to an AI systemUser asks chatbot a question
ResponseOutgoing response from AIChatbot’s answer
SignalPassive event or notificationSystem log, tool call

Review Task States

Review tasks (created by human intervention nodes) can be:
StateDescription
PendingWaiting for human decision
ProcessingBeing worked on
ApprovedHuman approved the request
RejectedHuman declined the request
Change RequestedHuman wants modifications

Integration Options

Velatir integrates into your existing AI workflows through:
  • Direct API - REST endpoints for custom integrations
  • SDKs - TypeScript/JavaScript and Python libraries
  • n8n Node - Official node for workflow automation
  • LangChain - Middleware for LangChain applications
  • MCP Server - Model Context Protocol for AI agents
  • Browser Extension - Monitor LLM interactions in browsers