Every week, three companies sent Seulgi's team pieces of the same picture. The numbers came in by phone or message, and someone on the team typed them into a raw sheet. A separate set of master tabs was supposed to pull those numbers in automatically, but the links broke often, and each time the number had to be retyped a second time, directly into the master. Phone confirmations went straight into the cell too, with no record of who'd confirmed what.
Packaging design changed every few months, so multiple versions of the same box were in stock at once. Each raw material could take six or seven rows just to show every packaging version in stock, each with its own run of weekly columns stretching back over a month, spread across tabs that didn't share a layout. Finding out what actually needed a phone call meant reading through all of it by eye and remembering what mattered.
The formulas weren't reliable either. One shipment line for the sushi nori product read $8,736. It should have read $81,984: a formula had pointed at the wrong column, and nothing on the sheet showed that anything was wrong.
What forced the issue was monthly shipments that went out short. The supplier didn't prioritize an order unless someone asked directly. Packaging shortages were buried in one tab among many, on a plant-by-material-by-week grid. By the time anyone noticed, there wasn't time left to order more, and orders to retailers shipped incomplete because no one could hold everything in view at once.
"It was all there somewhere," Seulgi says. "Finding it before it became a problem was the hard part."
The deeper issue wasn't that the spreadsheets lacked the information. It was that finding it meant paging through rows and columns by hand across tabs that didn't line up with each other. That kind of manual digging is exactly where mistakes get in, the kind nobody catches until it's already a missed shipment.
One list, sorted by what needs to happen. What's running short and needs ordering, what's missing information, who needs a call, and any other flag get pulled into a single ranked list, instead of sitting scattered across tabs someone had to cross-reference by eye and remember to check.
Every number opens to its source. Click any figure on the dashboard and it shows the exact report, week, and row it came from, plus who changed it and when.
Easy transition from spreadsheets. Paste a block straight from Excel and it lands on the right rows. Upload an entire existing workbook and it's read tab by tab, matched by column header, so nothing had to be retyped to get started, then or later.
The plants and the supplier enter their own numbers now. Instead of the team retyping what came in by phone or message, production counts, shipments, and packaging counts get entered directly, through screens the supplier can open with a single tap and no account required.
Packaging is tracked version by version, not just by total. Each packaging design version gets drawn down in the order it arrived, so a shortage in the specific version that's actually usable shows up even when the total count on hand looks fine.
They didn't build automatic reordering or flagging into this first version. The system captures how Seulgi's team actually works before it tries to make any of those calls itself. Balancing orders across multiple packaging vendors week to week also stayed manual. It's a negotiation, not a data problem, and it didn't fit the system well.
The dashboard went live first, about six weeks in, since it didn't depend on the rest of the build. When the full switch came, moving every existing sheet over took about thirty minutes, since the import screen read what they already had instead of asking anyone to start over.
The data reconciliation that used to take about eight hours weekly now takes ten minutes. The plants and the supplier enter their own numbers directly, so there's nothing left to retype and no broken link between tabs to fix before Seulgi can trust what she's looking at.
Mistakes that used to slip through get caught before they matter. Packaging film shared across products used to get counted once for every product that used it, so on-hand totals looked close to double what was actually there. The new system counts it once, at the material level, not the product level.
Finding what needed action stopped depending on memory and manual work. Scanning several tabs by eye for what needed a call used to eat up every morning. The ranked list now takes about five minutes to work through.
The switch itself cost nobody a day of downtime. Every sheet already in use, and every update sent afterward, came in through the same import screen. The full move took about thirty minutes.
Balancing orders across multiple packaging vendors is still done by hand, and a system that recommends reorders on its own is planned but not built yet.
"Going through sheets used to take most of the day," Seulgi says. "Now it's ten minutes, and I know exactly who I'm calling."
A newsroom's international desk used to lose the first hour of every morning just figuring out what was new. Now a tool does that overnight, and flags which stories the desk is likely to run.
Every operation has documentation that stopped being true at some point. This is a method for finding that point: extracting entities from records, building a graph of what should be, and reading the divergences against what is.
The pilot worked. A year later nothing is in production. The failure is almost never model quality. It is data nobody owns, permissions nobody granted, and a workflow nobody changed.
