INTERACTIVE SANDBOX

Orchestrate compliance with governed AI.

TIZNO connects visual rules, policy intelligence, and agentic execution, demonstrated here with synthetic data only. No real data, no network calls.

SCROLL
Live Demo

Experience the Orchestrator.

TIZNO bridges visual rule mapping with agentic AI execution. The guided scenario opens first, a mission-control walkthrough of a single compliance run, then use the tabs to explore workflows or the interactive sandbox. All synthetic data, no network calls.

Live scenario · read-only demonstration

Enhanced Due Diligence, Meridian Holdings Ltd · Risk Tier HIGH · Cross-border trigger

Synthetic · no network
VaultAtlasCRMOrchestrator
ORCHESTRATOR_STANDBY

Awaiting inbound signal from core banking read-only feed…

Workflows

Visual workflows to governed execution

Map rules without code, constrain what the Orchestrator may do, then ship every outcome through Action Cards and human approval. Use the Workflows tab for the drag-and-drop canvas preview. It is synthetic, with no live data and no outbound calls.

[Visual: Workflow Node Canvas]

Visual workflow canvas

Compliance and operations teams draw High-Risk EDD flows, Account Restriction flows, and other regulatory paths on a drag-and-drop canvas. Triggers, branches, and terminal actions are wired visually, no deployment pipeline or hand-authored scripts.

This layer is the contract for automation: it defines which events may start a run and which tools the stack may call. Guardrails are fixed before the Orchestrator executes, not inferred at runtime.

  • 01Trigger graph: eligible entry conditions (risk tier, PEP flags, jurisdiction, product).
  • 02Tool allow-list: Vault scopes, vector corpora, and draft generators the agent may invoke.
  • 03Versioned workflow definition: immutable reference the Orchestrator must satisfy per run.
ReAct

[Interactive Tour: Step 2 · Orchestrator trace]

Reasoning engine · Orchestrator · Agentic RAG

The assistant does not idle in open-ended chat. It follows a ReAct-style loop: read structured state, call approved retrieval tools, observe results, then produce a bounded artifact with explicit citations.

Example: the stack profiles a high-risk corporate client, finds gaps in KYC fields sourced from the bank's CRM, runs semantic vector search across the 2025 Gibraltar National Risk Assessment and public GFSC manuals, and drafts an Enhanced Due Diligence memo with exact page and section citations.

Vault read: tenant-scoped CRM / core payloads, RBAC-enforced columns and entities.
Atlas RAG: pgvector retrieval over NRA + GFSC embeddings, ranked by relevance scores.
Orchestrator synthesis: EDD memo body plus inline source anchors tied to each claim.

[Visual: Action Card · Human-in-the-loop gate]

Execution & governance · Action Cards

The model never receives uncontrolled write access to the CRM. It emits Action Cards, examples include "Drafted email to client" or "Proposed account freeze", each bundling the proposed operation, affected identifiers, and policy rationale.

A human compliance analyst must choose Approve or Reject. Rejection feeds structured feedback back into the Orchestrator for another constrained pass. Approval dispatches the action and records an immutable, timestamped execution trace suitable for internal and regulatory audit.

  • Action Card payload: operation type, client IDs, citations, risk summary, draft content.
  • Human gate: binary approve / reject with optional rejection reason for model steering.
  • Execution log: API fetches, tool calls, reasoning steps, and sign-off in machine-readable form.

[Visual: approval chrome]

Pending Action Card

Proposed account restriction

Awaiting compliance analyst

ApproveReject

Decorative preview, interactive controls live in the Orchestrator steps tab.

Guided scenario

Infrastructure in motion, not a tooltip tour

This sequence mirrors a real high-risk run: core banking trigger, Atlas clause lock, read-only CRM pull, Orchestrator synthesis, and Enclave holding an Action Card until a human approves. When you are ready, switch to Interactive steps or use "Run it yourself" after the final frame.

All data shown is synthetic and generated for demonstration purposes only. No real client data is used.