Meridian

How a customs declaration becomes an opportunity

Two halves that meet only through files on disk: a Python pipeline that turns raw extracts into typed, flagged rows, and a Next.js app that reads its committed output. The app never runs the pipeline and never reads its working files — only the payloads in public/. That is why Vercel can build this without Python, pandas, or the 540 MB intermediate CSV.

Every number on these pages is measured from the artifacts, not typed into them. Sizes are read from disk on the server; counts come from a payload the page already loads. This exists because the numbers that were typed — all of them, in README.md — had drifted by two to three times.

The evidence/ row is one representative file. There are 40 of them, one per covered account, totalling 19.12 MB — fetched one account at a time, never as a set, which is why the directory can be larger than everything else combined without costing a page load. See what each payload holds and what breaks when it is stale →