How I Framed It

I mapped the decision each department makes daily and worked backwards to the data required.

The buying team's core decision is: what do I reorder today, and how much? That requires sell-through velocity, current stock, lead time, and safety buffer — all in one place, ranked by urgency. The marketing team's core question is: what are customers looking for that we're not serving well? That requires search gap analysis, CTR by query, and trending terms. Leadership needs a 10-second health check: SKUs at risk, revenue at risk, OOS count.

Rather than building one monolithic dashboard, I built the product as a set of purpose-specific views — each one answering a department's primary question — backed by a shared data layer that pulls from Shopify, Algolia, Google Search Console, and the WMS on a 15-minute cache refresh.

The design principle throughout: zero manual input. Every number is live or cached from a source of truth. The dashboard surfaces what the data says, not what someone chose to enter.

Key Decisions

Build internally, not buy a third-party inventory tool

Context: There are established Shopify inventory tools (Stocky, Inventory Planner) — cheaper in the short term. But the buying team's workflow was deeply specific: Bewazir consignment catalog, WMS reconciliation, Arabic-language context. No off-the-shelf tool connected Shopify, WMS, Algolia, and Google Search Console in one place.

Choice: Custom internal build on Next.js. Zero per-SKU fees, no data leaving the ecosystem, full control over the UX. The internal build became the only place that could honestly answer questions like 'which Bewazir-consigned SKUs are about to go OOS and which supplier do I call?'

Trade-off: Significant build time and ongoing maintenance responsibility. Every new data source requires custom integration work.

15-minute cache over real-time streaming

Context: Real-time streaming from Shopify's API is technically possible but architecturally complex and expensive at scale. Options ranged from webhook-driven streaming to daily batch refresh.

Choice: 15-minute cron-warmed cache. For buying decisions, 15-minute data is effectively real-time — no one is making reorder decisions faster than that. The cache meant every page load was instant (pre-warmed), reducing API quota usage and eliminating cold-start latency for the team.

Trade-off: A 15-minute blind spot on flash OOS events. Mitigated by a separate hourly WMS sync that reconciles Shopify inventory against warehouse quantities and flags desyncs.

Dead Stock Triage as a dedicated full-screen mode

Context: Dead stock (inventory with zero sales in 30/60/90 days) was the buying team's most painful unresolved problem — worth tens of thousands of SAR in tied-up capital. Surfacing it as a filter on the main products table had been the existing approach and it wasn't moving the number.

Choice: Dedicated full-screen triage mode where each dead stock item is reviewed one by one — with an AI-generated action suggestion (promote, bundle, liquidate, or remove) specific to that SKU's category, value, and supplier. The one-at-a-time format forces decisions rather than enabling indefinite deferral.

Trade-off: A heavier build than a simple filter. The AI suggestion quality depends on metafield coverage — SKUs with thin data get generic recommendations.

AI features as workflow accelerators, not novelty

Context: Every dashboard in 2026 has an AI feature. The risk is building AI that gets used once and ignored. A general chat over the data is easy to build and hard to make load-bearing.

Choice: Targeted AI at specific bottlenecks. The Daily Buying Brief surfaces every morning as a 120-word summary of inventory health with three urgent actions. Reorder email drafts are AI-generated on demand because writing supplier emails in English and Arabic was a real daily friction point. Dead stock suggestions are AI-generated because the action (promote/bundle/liquidate) requires judgment the data alone can't provide.

Trade-off: More complex to build and maintain than a general chat. Each AI feature needs a prompt, a data schema, and a quality bar — and all three require iteration.