Online storesshipped
Weft
Operations layer for global e-commerce: keeps stock and orders correct across a headless Medusa store and Shopify, pushes paid orders to Xero without duplicates, and ships an AI assistant that is measured and is not allowed to invent a price.
- Timeline
- Completed
- Role
- Full Stack Developer
- Team
- 1 person
- Hub tests 139 passing
- Assistant eval 59/60, holdout 23/24
- Demo 0:49

Overview
Weft is a portfolio build of the part of e-commerce that breaks quietly: keeping stock and orders in agreement across systems. The demo store is Halden Moto Parts, a fictional motorcycle parts shop where fitment ("does this fit my bike?") is the product. A hub keeps a stock and order ledger, talks to Medusa, Shopify and Xero, and exposes a console for exceptions and reconciliation, plus a customer assistant.
The Challenge
- Several systems can disagree about stock and orders, and retries can create duplicates: a double invoice, an oversold item, a failure replayed forever.
- The assistant also must never state a price or stock figure it did not get from the system.
The Solution
- Stock is a ledger of deltas with unique event ids, and orders reserve stock in one transaction with ordered row locks.
- Outbound work goes through a transactional outbox with leases, backoff and dead-lettering into an Exceptions queue, and idempotency keys are scoped to each attempt.
- Reconciliation compares Weft with Shopify and Xero and only auto-fixes what a machine can safely fix.
- For the assistant, code decides when tools run, answers about price and stock are built from tool results, and a guard rejects any figure that no tool or cited passage backs.
