What is this?
You're looking at the Leftbank Art Dust Agent System — a simulation of how AI agents handle the customer-facing work of an art wholesale + ecommerce business. Everything here is synthetic: no real customer or order data, no real money moving.
Three things are happening continuously
- AI personas (24 of them across 6 archetypes — boutique buyers, designer-trade buyers, big-box buyers, online-only retailers, direct consumers, and artists) act like real customers. They place orders, ask questions, request refunds, complain.
- AI agents (Triage + Customer Service + Order Fulfillment) respond. Triage classifies each inbound message and routes it; CS and OF read the message, look up the relevant order or product in the mock storefront, draft a reply, and (if low-risk) auto-send it.
- A human reviewer approves anything money-touching (refunds, replacements, complaints) via the Approval Queue.
Where to look
- Overview — today's stats at a glance.
- Activity — live plain-English feed of every action.
- Chat — type a message yourself and watch the agents handle it in real time. Pre-loaded examples save you typing.
- Personas — click a persona to see who they are and trigger any of their actions on demand.
- Shop — browse the synthetic catalog (200 SKUs).
- Mailpit — see actual emails flowing in and out.
- Approval Queue — pending agent actions waiting on a human.
- Grafana — engineering-flavored metrics.
What's not happening
- No real customers are being emailed. The "Mailpit" surface is both the inbox AND the outbox in simulation mode.
- No real money is moving. Orders, refunds, and replacements all live in a synthetic database.
- No real catalog. The 200 products are deterministically generated.
Why it matters
Once Leftbank goes to production (Phase 3+), the same agents handle real customers — Triage routing, CS drafting, OF flagging stuck orders. The simulation lets us measure quality, cost, and approval-gate behavior at realistic volume before connecting to real systems.