A label system a warehouse floor runs on
Built July 2025, in daily use
I built a browser-based warehouse tool that turns inconsistent CSV and Excel exports into scannable 4x6 thermal labels. Operators scan a barcode, resolve the product, assign a batch, and print product, date, allergen and box data on a Zebra ZT610. I built it while working in warehouse logistics, and it has generated 144,000+ labels since print tracking began in August 2026, about 2,400 on a working day.
What it replaced
The warehouse relabels product before it ships. Before this, it was done by hand: print an Excel template, fill it out with a pen, then write the details on each box with a Sharpie. That was slow and easy to get wrong, and a wrong label on a box is a real operational problem. Now an operator scans a barcode and pulls a finished label off the printer.
What it has to survive
The data comes from a state track-and-trace system and a transfer tool as exports in different shapes, with the header in different places. The importer finds columns by reading the header text, so a slightly changed report still loads.
Two facilities print different orientations, on a floor where the network is not guaranteed. One renderer handles both layouts. The product catalog falls back from the database to a verified local copy, and standard printing still has to pass its own freshness checks, so a dropped connection cannot quietly put stale data on a label. A product it cannot recognize gets logged for an admin to remap instead of being dropped.
What broke
The scanners. Many of the handhelds strip the first character of whatever they read, so a normal barcode came back missing a digit and every lookup failed. Each label now encodes a tilde in front of the value, so the scanner strips the tilde. You only learn that one by standing next to the hardware.
My own fix for Excel. Long barcodes passed through Excel can come back as 7.75984E+15. My first fix rounded that back into a full number, which looked right and was wrong: those digits are gone, so it made up a barcode that could then be printed on a real box. Now that value is a hard error on the row, and only the unsafe row is quarantined, so one bad cell does not block the whole report.
My own checks, again. I added warnings for odd-looking columns and based them on printed labels. On the first real export they fired on normal rows, because most rows in a report are bulk stock that never gets a label. They now reject only values that cannot belong in a column. I tested the next version on a real export first.
The browser. If it drops the drawing canvas partway through a job, the app now stops the job instead of printing a black page.
Measured use
144,000+ labels since tracking began in August 2026. Two facilities. Named operator accounts, with every print logged under the person who ran it. A review step before every print, on runs of up to 5,000 labels in a single PDF. The current version is 7.8.
Built with React, Vite, Supabase and Postgres, Canvas and jsPDF. Unit, real-browser label and migration tests run in CI.
Scope. These are operational labels, not certified compliance labels. The allergen flag is a keyword check that prompts the operator, not an ingredient database. It is an internal tool for a fixed set of warehouse users, not a public app. The code is private because it runs on a real employer's inventory.