Product Lab case study · Granite Labs

Brand
Product
Engineering
AI

Sourcing research

The Product Lab review screen for a Wella hair colour listing: supplier row, matched Amazon listings by pack size, checks and unit economics, tilted on a dark plane of neighbouring products.

Case study · 2026

The sourcing researcher you never have to hire. Product Lab checks every row of a supplier price list against Amazon and brings back what qualifies.

What we did

Brand identity
Product design
Web app and website
API and data pipeline
Model training

Long story short

Granite Labs designed and built Product Lab end to end: the brand, the app, the website, the API and the model training pipeline.

Highlights

Listings checked in the US, UK and EU
Images, pack and variant before margin
One keystroke when a pack is unclear
Forty rules, one master per component

For

Amazon resellers sourcing from supplier price lists

Markets

United States, United Kingdom and the EU marketplaces

The Product Lab box mark drawn on its construction grid, with measurements in blue.

The listings card on yellow, annotated: each colour has one meaning, from qualifies in green to your call in orange.

The challenge

Up to 50,000 rows

A supplier price list can run to thousands of rows. Each one has to be matched to its Amazon listings, checked for pack and variant, and priced against fees before anyone decides what to buy.

The product had to do that work in the background and stay honest about doubt. When a listing's title says six and its photo shows three, a person makes the call.

AI search: 38,520listings from 9,630 searches
AI screen: 4,280could be your product
AI photo check: 1,912matched, every pack countedmatched, units in every pack counted
Priced on your costs: 196listings worth buying

One price list, four passes

Search, screen, photo check, pricing

The checks table for one match: six points from title to pack, each marked before the listing is priced, with the pack row outlined.

The product plane

Fly through review. Nothing ever loads.

Review by keyboard

One keystroke per answer

The approach

One way to draw each thing

The screens had drifted into two to five versions of each element. We wrote forty hard rules and gave every component one master, so the app, the website and the sign-in pages speak one language.

Paper tokens are the code tokens, and a lint of eight checks runs before any board loses its review tag. Yellow never marks a problem. It marks the brand, where you are, and a run that is still going.

FFF6D6#FFF6D6
FFE59A#FFE59A
FFD45C#FFD45C
F2B93B#F2B93B
7A5600#7A5600

One yellow, five steps

Brand, place, live runs. Never a status.

Hanken Grotesk Bold specimen: a large Aa.

Paper Mono specimen on ink: an ASIN and a barcode number, set in yellow and white.

Three moods of the box mark on cream: hello, checking and done.

The build

Five apps, one repo

One Turborepo holds five apps: a NestJS API, the TanStack Start app, the Next.js website, a generator that builds training data, and a Python trainer that fine-tunes small language models on Modal.

Runs queue on Redis and carry on when the tab is closed. Results land in Postgres through Drizzle, and files in Cloudflare R2. Every screen ships from the same Paper tokens.