Building Scalable E-commerce Platforms: Lessons from Jou Sun

1 Start with the constraints, not the stack

The best architecture I've shipped started as a checklist of constraints: dispatch windows, inventory integrity, supplier lead times, refunds, and the reality that operations staff can't be forced into "developer workflows". Once those are explicit, the stack becomes a consequence.

2 Model the supply chain as a first-class system

When grocery delivery works, it's because the supply chain logic is aligned with what customers see. That means inventory isn't just a number — it's state. "Reserved", "picked", "substituted", and "failed to fulfill" are all different, and each one needs an audit trail.

• Idempotency for order creation and payment callbacks.
• Event logs for inventory mutations and fulfillment steps.
• Operational UX that makes the correct action the easiest action.

3 Build B2B tooling like a product

The internal admin isn't a "back office" — it's where the business actually lives. When the admin becomes slow or confusing, the business loses time, then consistency, then margin.

Two moves made the largest difference:

1. We treated admin screens as workflows, not pages — each screen had a defined "finish line".
2. We used guardrails: defaults, validation, and system-generated suggestions instead of requiring manual precision.

4 Choose integration boundaries you can change later

Commerce platforms evolve continuously: payments change, delivery options expand, suppliers rotate. The key is to introduce boundaries where you expect churn. In practice, this meant a stable internal order model and isolated adapters for payments, fulfillment, and notifications.

Order
  - customer
  - basket
  - pricing
  - inventory_state
  - fulfillment_state
  - payment_state

Adapters
  - PaymentProvider
  - FulfillmentProvider
  - NotificationProvider

5 Performance is a product feature

Fast pages make customers trust the platform; fast admin tools make teams trust the system. The best ROI came from boring, repeatable work:

• Cache expensive reads (but never at the cost of correctness).
• Precompute "next actions" for operational dashboards.
• Measure slow queries in production and fix the top 10.