Public Architecture

Not a chatbot wrapper.
An operating system.

Most "AI marketing tools" are 5-10 GPT prompts in a UI. Zbon is an operating system: a shared memory, an orchestration layer, an evaluation framework, a guardrail engine, and a human-in-the-loop approval system that bind 9 specialist agents into one coherent revenue engine. This page explains how it works — publicly and honestly.

The four-layer architecture

Every action in Zbon passes through these four layers, in order.

LAYER 01

Orchestration Layer (Zbon Core)

The control plane that decides which agent runs, when, and in what order. When a cart is abandoned, Orchestration routes the event to Customer Memory → Content Director → Email Manager — in that order, with the correct context passed between them.

  • Event-driven dispatcher (webhook + cron + manual triggers)
  • Dependency graph between agents (Strategy → Content → Media)
  • Retry logic with exponential backoff on agent failures
  • Cost ceiling enforcement — no runaway LLM bills
LAYER 02

Memory Layer (5 stores)

Zbon remembers. Most tools start fresh every prompt. Zbon has 5 distinct memory stores, each with a defined schema, retention policy, and update cadence. Agents query these — they do not re-derive context every call.

  • Brand Memory — voice, tone, banned phrases, taxonomy (read by every content agent)
  • Customer Memory — identity graph, lifecycle stage, LTV, churn risk
  • Performance Memory — winning creatives, seasonal patterns, attribution
  • Revenue Memory — channel mix, recovery rates, financial truth
  • Campaign Memory — every prior campaign with outcome (agents reference, not reinvent)
LAYER 03

Agent Layer (9 specialists)

Each agent is a focused unit with a defined responsibility, defined inputs, defined outputs, and a defined KPI it is graded against. Not a single mega-prompt. Not interchangeable. Each runs on the model best-suited for its task (e.g. Content on Claude, Data on Gemini).

  • Strategy Director · Content Director · Media Buyer · SEO Director
  • Email Manager · Recovery Engine · Data Analyst · Research Engine · Social Manager
  • Each agent has a public KPI it is evaluated against weekly
  • Multi-provider model routing (Claude / Gemini / Groq / OpenAI fallback)
LAYER 04

Governance Layer

The safety rails. Nothing executes outside this layer. Every output is checked against Policy Center rules before it goes anywhere. Every action is logged. Every action is reversible for 24 hours.

  • Policy Center — spend caps, banned phrases, escalation thresholds
  • Approval Inbox — humans approve before risky actions run
  • Decision Log — every agent decision timestamped, reasoning visible
  • Rollback — undo any agent action within 24 hours

You choose how much autonomy

Three trust tiers. You can change them per agent, anytime.

Advisor

Zbon drafts everything. Nothing executes without your explicit approval.

Best for: first 30 days, regulated industries, high-stakes ad spend.

Co-pilot

Low-risk actions (< $200) auto-execute. High-risk actions wait in Approval Inbox.

Best for: experienced merchants comfortable with the agents.

Autopilot

All actions auto-execute within Policy Center limits. You audit weekly.

Best for: high-volume stores trusting the system end-to-end.

Transparent stack

We do not pretend to have built our own LLM. Here is what powers Zbon.

AI models
Anthropic Claude + Google Gemini + Groq + OpenAI (fallback chain)
Frontend
Next.js 14 on Vercel (Edge runtime)
Backend
Node.js + Express on Render
Database
Supabase (PostgreSQL + pgvector)
Cache & queue
Upstash Redis
Payments
Stripe (PCI-DSS Level 1)
Email
Resend (verified noreply@zbon.net)
Vector memory
pgvector for semantic search across memory layers

Honest limits

We are in closed beta with founding merchants. Some things we have not yet built or formalized:

  • No SOC 2 / ISO 27001 yet (audit starts after first 50 merchants).
  • No official PDPL / ZATCA / NCA certifications yet (architecture is ready; certifications follow merchant volume).
  • No production Mada / STC Pay / Tabby / Tamara integrations yet (Stripe handles cards now).
  • No data residency in KSA yet (we use US/EU regions via Vercel/Supabase).
  • No SLA contracts yet (best-effort 99.5% uptime currently).

We add these as we scale. Founding merchants get direct founder access to influence the priority list.

Want to see this running on your store?

14-day free trial. No credit card. Cancel anytime. The architecture above is what runs the moment you connect.

Connect my store