Coco Ding

Some of the underlying business data and implementation details are proprietary. This case study focuses on the architecture, problem-solving approach and technical decisions rather than confidential commercial information.

The problem

The existing vendor connector could only work with one Shopify store at a time. That created an awkward ceiling for a business operating across four markets.

The data existed. The API existed. The interface between the two was the problem.

The question

What if the four stores could be treated as one system?

What I built

A multi-store Shopify data layer covering 4 stores, 13 tools, 3,463 lines. The system includes:

• OAuth client credentials
• automatic token refresh
• Keychain-managed secrets
• concurrent fan-out
• per-store failure isolation
• read-only defaults
• explicit write access per store

The interesting part

The hardest part wasn't calling the API. It was designing the system so that one store failing didn't bring down the other three. And making sure that "read-only" actually meant read-only unless a workflow explicitly required otherwise.

The principle

Don't accept a tool's ceiling as a system's ceiling.

What changed

Instead of repeatedly working around a one-store-at-a-time connector, the underlying Shopify data could be accessed across four markets through one reusable layer. That made a different class of analysis and automation possible.

What I learned

A limitation in the interface isn't necessarily a limitation in the underlying system. Sometimes the right answer is to go one layer deeper.

This is a reconstructed representation of a real operational problem. Commercial values and business-specific data have been omitted or recreated where necessary.

The problem

Every week, stock management involved downloading data, cleaning spreadsheets, moving information between teams, checking hundreds of SKUs and working through competing stock requirements. The process took around three manual hours.

More importantly, it created opportunities for human error and made a decision that should have been systematic feel surprisingly subjective.

The question

Could the weekly stock decision become a system rather than a spreadsheet exercise?

The inputs

• Current inventory. What is available now?
• Demand velocity. What has been selling?
• Inbound inventory. What is actually coming?
• Shipment timing. When is it likely to arrive?
• Wholesale commitments. What stock is already spoken for?
• Stock buffers. What needs to remain protected?
• Recent stockouts. What demand may already be going unserved?

The system

ShopifyCurrent stockSell-throughDemand coverageInbound shipmentsWholesale constraintsStock bufferAvailability gapsDecision

Where AI fits

AI is used as part of the analytical and decision workflow. It helps interpret data, develop and maintain the workflow, and surface patterns across multiple inputs. Deterministic logic remains responsible for critical validation and calculations.

The important bit

The system does not assume that a number is correct because it looks reasonable. It reads spreadsheet columns by formula signature rather than relying only on column headers. If the expected structure isn't there, the workflow stops.

Why that matters

A system that says "I can't safely answer this." is more useful than a system that confidently produces the wrong number.

What it found

The broader workflow reduced approximately 200 hours of annual manual work.

What I learned

The difficult part wasn't automating the spreadsheet. It was deciding what information actually belonged in the decision.

The observation

A new collection was receiving traffic from advertising. Visitors were landing directly on the collection page. The page, however, gave them very little beyond a collection title and product cards.

What I saw

Session recordings showed that the friction was happening early. Customers had arrived with intent. But the page required them to take additional steps to understand the products.

The question

Where are we asking customers to work harder than they should have to?

The evidence

• heatmaps
• session recordings
• Shopify performance data
• funnel behaviour
• device-level analysis

What I did

I translated the findings into prioritised UX and merchandising improvements. I also led development with the external web agency, managing a backlog of 20+ website improvements and bug fixes. Each brief included:

• commercial rationale
• expected customer impact
• priority
• before / after UX references
• implementation requirements

Result

Within one month in NZ:

The insight

Sometimes conversion optimisation isn't about persuading someone harder. It's about removing the unnecessary work between "I'm interested." and "I understand this product."

26,825 accounts

A customer-tagging gap affected 26,825 accounts. The investigation went beyond the vendor's proposed explanation and overturned the suggested fix.

2,398 orders

An audit of 2,398 authorise-to-capture orders identified the underlying payment-flow issue and stopped a refund being issued for a problem that wasn't actually a refund fault.

Variant-level collections

Shopify's standard collection interface couldn't describe the required variant-level logic. The API could. So I built the collection at the level the business actually needed.

The pattern

When the interface can't answer the question, go underneath it.

The question

Could the data tell us something before the next product decision was made?

What I analysed

Three years of cross-market sales and customer data. Including:

• product velocity
• best-seller performance
• repeat-purchase behaviour
• first-to-second-product relationships
• customer-spend thresholds
• model performance
• colour performance

What I was trying to understand

Does the first product a customer buys tell us anything about the second? Does customer spend relate to the likelihood of returning? Are some products good at generating a second purchase even if they aren't the initial best seller?

Why it mattered

Limited-edition planning had historically relied heavily on subjective judgement. The analysis introduced evidence from historical sales, product performance, colour trends, audience characteristics and previous collaboration performance into future product and marketing discussions.

The broader point

The interesting question wasn't "What sold the most?" It was "What can we learn from the pattern behind what sold?"