Take in
Retailer PO reports and shipping questionnaires are uploaded, previewed and confirmed by a person before anything is written.
Mordéa Labs · Sydney, Australia
Every retail order, on one thread.
Mordéa Labs builds Mordéa Thread, the AI operating system for home-goods brands that sell to national retailers.
From the retailer's purchase order to the factory floor, the container and the invoice: one record, run by people and AI agents under the same rules.
Purchase orderFactory orderDeposit confirmedIn productionQC passedShippedInvoicedBuyer paidReconciled
The problem
One order. Six places. One wrong attachment.
A national retailer sends a purchase order. It becomes factory orders, deposits, QC dates, containers, invoices and payment terms. Each lives somewhere different: a spreadsheet, a PDF, a buyer portal, a freight booking, an email thread, a chat with the factory.
Staff copy the same numbers from one to the next, and the risk that matters most is quiet: a factory receives a file that still carries the retailer's price.
- Spreadsheet
- PDF purchase order
- Buyer portal
- Freight booking
- Email thread
- Factory chat
Thread keeps them as one record.
Mordéa Thread
From purchase order to reconciled, on one timeline.
Thread takes in the purchase order, splits it into factory orders, tracks production and shipments, and closes the loop on invoices and payments. Screens below are the live product with demo data. Each deployment carries the operator's own brand.
Split
Purchase orders, factory orders, allocation and shipments on one timeline, from PO received to reconciled.
Ship
Containers, departure dates and quantities shipped against each factory order, so what is left to ship is never a guess.
Close
Four roles, each with its own pages and fields; every change lands in an activity log.
Factory portal
The factory sees its cost. Never the retailer's price.
Each factory signs in to its own orders and sees its own agreed cost. Retail prices are removed on the server, not hidden on the screen.
Retailer purchase order · seen by the brand
PO SR-2026-0917-01
- SKU
- SB-NAT-01
- Product
- Storage basket · Natural
- Qty
- 1,200
- Retail unit price
- AUD 8.00
Factory order · seen by Demo Factory
FO-000002
- SKU
- SB-NAT-01
- Product
- Storage basket · Natural
- Qty
- 1,000
- Retail unit price
- Removed on the server
- Agreed unit cost
- AUD 5.00
Demo data
Agent harness
AI agents that work under the same rules as staff.
AI agents work through the same typed tools and permissions as staff; every write needs an approval and every action is logged. Agents hold no credentials.
Skills describe the work. Tools hold the keys. The record decides.
- 01
Surfaces
The admin console today; Slack threads next. People ask, supply missing details and approve.
- 02
Orchestration
Tasks, runs, waits, resumes and budgets.
- 03
Policy
Each channel maps to a tier, an approver, a model and a budget.
- 04
Tools
The only layer that holds business credentials. It filters fields, forms proposals, deduplicates and returns receipts.
- 05
System of record
Thread's own rules, states and transactional writes decide the final answer.
14typed tools over MCP. Every call carries a run reference and the person it acts for; tokens are issued and revoked by the owner.
SydneyThe harness is deployed in the operator's own AWS account in Sydney.
0credentials held by agents. Writes pass the same pre-write checks as a person's.
Illustrative ledger · demo data
run 7f3aget_po SR-2026-0917-01readacting for Owner Demorun 7f3alist_fos SR-2026-0917-01read1,200 unallocated → 200run 7f3bproposal: convert remainder to FOawaiting approvalapprover: Owner Demorun 7f3bconvert_po_to_foapproved · writtenreceipt FO-000003run 7f3clist_events FO-000003readread back: matches
The first agent: Orders In development
Status, explained
Where a purchase order stands, what is still unallocated, and exactly what is missing before it can become a factory order.
Follow-up
Finds the orders that need a person, opens a task, and waits for the answer without asking twice.
PO email intake
Reads the email and attachment, matches buyer and products, checks the price source and proposes a candidate order.
Factory orders
Collects factory, quantities and terms into a proposal; a person approves; the agent writes and verifies the result.
Status
What runs today, and what is next.
| Module | What it does | Status |
|---|---|---|
| Order record | Purchase orders, factory orders, allocation and shipments on one timeline, from PO received to reconciled. | Live |
| Spreadsheet intake | Retailer PO reports and shipping questionnaires are uploaded, previewed and confirmed by a person before anything is written. | Live |
| Factory portal | Each factory signs in to its own orders and sees its own agreed cost. Retail prices are removed on the server, not hidden on the screen. | Live |
| Roles and audit | Four roles, each with its own pages and fields; every change lands in an activity log. | Live |
| Agent tools | 14 typed tools over MCP. Every call carries a run reference and the person it acts for; tokens are issued and revoked by the owner. | Live |
| Agent harness | Orchestration, approvals, budgets and a ledger, deployed in the operator's own AWS account in Sydney; agents do not yet act on live orders. Agents hold no credentials. | In development |
| Orders agent | Explains where each order stands, chases what is missing, turns PO emails into candidate orders and proposes factory orders for approval. | In development |
| Inbox capture | Orders arriving by email and chat become candidates on the record. | Next |
Next: models on NVIDIA, in Australia
Read every purchase order without the data leaving Australia.
Retail purchase orders carry commercial terms that brands do not want to send offshore. Next: NVIDIA NIM with Nemotron models on NVIDIA GPUs in Sydney, so purchase orders are read without commercial data leaving Australia; NeMo Retriever for document extraction; NeMo Guardrails to hold agents to supplier-safe rules.
NVIDIA NIM · Nemotron
Serve the models that read purchase orders and draft candidate orders on NVIDIA GPUs in AWS Sydney, next to the record they write to.
NVIDIA NeMo Retriever
Extract line items, quantities and dates from PDF purchase orders and retailer spreadsheets into Thread's intake preview.
NVIDIA NeMo Guardrails
A second rail at the model boundary: agents never put a retail price in front of a factory and never write without an approval.
First deployment
Live at Mordéa Home.
Thread is live at Mordéa Home, a sister company and its first deployment. Mordéa Home (ABN 65 606 115 762) designs storage and homewares for national retailers in Australia and North America; its retail orders, factory orders and shipments run on Thread.
Company
Mordéa Labs Pty Ltd, Sydney.
Mordéa Labs Pty Ltd builds and runs Mordéa Thread. Privately held. Based in Sydney, New South Wales, Australia.
Mordéa Labs builds Mordéa Thread, an AI operating system for home-goods brands that sell to national retailers. One retail order touches a purchase order, several factories, sea freight, invoices and payment terms, usually spread across spreadsheets, PDFs and email. Thread keeps it as one record: it takes in the purchase order, splits it into factory orders, tracks production and shipments, and gives each factory a portal that shows its own agreed cost and never the retailer's price. AI agents work through the same typed tools and permissions as staff; every write needs an approval and every action is logged. Thread is live at Mordéa Home, a sister company and its first deployment; the agent harness is deployed in AWS Sydney. Next: NVIDIA NIM with Nemotron models on NVIDIA GPUs in Sydney, so purchase orders are read without commercial data leaving Australia; NeMo Retriever for document extraction; NeMo Guardrails to hold agents to supplier-safe rules.
- Yiwei HuangFounder and DirectorLeads product and the brand side of every order flow Thread runs.
- Ruixuan LiaoDeveloperLeads engineering on Thread and the agent harness.
Contact
yiwei@mordeahome.com