↺ Back to builds

Internal Tool · Private

Inventory Showcase

What to buy. What to showcase. What to exit.

Astro 7.3 · Tailwind 4 · Cloudflare Workers + D1 + R2 · TypeScript engine (zero-IO) · GitHub Actions collectorsEvery piece captured with photo + ~10 structured tags; capital-at-risk computed per item, not feltDeterministic demand engine — scores carry confidence bands that scale with independent-source agreement; no ML, the weekly loop is the modelOne WhatsApp message a week in rupee language: buy now, exit list, capital locked, next 8 weeks

The Friction

Jewellery inventory is capital sitting on hooks, and a design-led brand makes its inventory decisions by taste alone — buy what looks right, show what feels right, and quietly accumulate designs that will never sell. Gold softens the error but doesn’t erase it: dead stock isn’t loss, it’s capital locked for months plus making charges never recovered. And at pre-launch, for this brand, inventory operations didn’t exist at all.

The Catalyst

“Nothing exists yet” turned out to be the permission: instead of bolting a tool onto a running store, the operating brain gets built first. The brand’s own kitchen architecture — starters, mains, chef’s table — became the demand taxonomy before any code did. And a single sentence became the charter for the whole system: what to buy, what to showcase, what to exit — always measured in rupees of capital at risk.

The Resolution

Three layers, one pipeline. The item ledger: every procured piece photographed, tagged, state-tracked, and valued at capital-at-risk. The intelligence layer (Tola): a fully deterministic engine scoring style archetypes against festival calendars, gold prices, and market signals — every piece inherits its archetype’s score with a confidence band. Then the customer-facing showcase: the storefront itself — homepage slots, menu order, drop layout — run as a digital merchandising calendar, sized in showcase slots, because shelf space and not rupees is the real constraint.

The Pipeline

Built honest before built clever: v0 runs entirely by hand — a sheet ledger, a curated festival calendar, a report template. v1 becomes code: capture PWA that survives field connectivity, collectors for marketplace velocity with a manual fallback flag, the zero-IO engine package, and the weekly report — one message, rupee language, no dashboards. The launch webhook is the milestone that matters: the brand’s own sell-through becomes the calibration anchor the moment real orders exist.

The Horizon

The same stack turns outward as a service line for the industry: city-level demand intelligence for traditional jewellers staring at a contemporary-assortment blind spot — sold as aggregates only, with the brand’s own item data never leaving the building. Discipline stays written in: sell capital-at-risk reduction and scenario bands, never point forecasts. And the storefront’s merchandising calendar becomes what a good engine always becomes — boring, because every important thing in it already computed itself.