23,658 products, two dollars of compute

About 23,658 products needed catalog cleanup. Categories were inconsistent, datasheets were missing and only 0.45% of items had an image. There was also a library of roughly 26,600 unclassified files. Running every field through a model would be expensive and hard to verify. Filenames weren't reliable enough to classify the documents.

I used rules and lookups for the straightforward work, which resolved about 60% of recategorization without model calls. Agents handle the remaining judgments; code controls provider calls, writes and budgets. A separate verifier checks outputs. People review disagreements and a random sample. The budget stops the process at 80%, and phase approvals go into an append-only record. Three review rounds found roughly 46 issues before production work. Document inspection also found six internal invoices with customer PII and one file with partial payment-card data in a ten-file sample. Every file now goes through a quarantine check.

The pilot recategorized 697 products for under $2 of compute. The independent verifier agreed on 98.1% of pairs and 100% of category values. The full program's cost is projected at roughly a thousandth of naive per-product prompting; that is a projection, not a completed-program result. ERP investigation also uncovered 134,058 inspection images in 31,250 folders. The larger program remains subject to phase approvals.