Document Types

What borrowers actually send us: documents split from uploads, classified by type — and how well the classifier keeps up

snapshot

Range

Documents per day, by type

Stacked by classified type · line = borrower uploads (right axis; each upload splits into one or more documents) · classification began May 6 — earlier days are all "unknown"

Classifier quality

Share of documents left "unknown" · dashed = trend · down is good — every unknown document is one a human may have to open

Type league table

Trailing 90 days · parse fail = file-level hard failures · dupes = auto-detected duplicate documents (detection began May 21)

Methodology. One row per document in files_file joined through packages_loanpackage; internal/test tenants excluded (same list as the Revenue Report). Nearly every document is a segment split from a borrower upload (source_file_id); the uploads line counts distinct source files per day. Type classification began 2026-05-06 — every document before then is "unknown", so the classifier-quality panel starts there. Duplicate detection began 2026-05-21 (duplicate_of_id). Types are lower-cased (`unknown` and `Unknown` merge); blank/null becomes "unclassified". Avg pages from start/end_page_index. Parse fail = file status 3. Weekends (Sat/Sun, Eastern) excluded from charts and KPIs; lines smoothed with a centered moving average (3-day at 30d, 5-day at 90d; none at 7d) while KPIs and tables stay raw. File names and file contents are never inlined (they can contain borrower details) — types, counts, and page stats only. Rolling 90 days · laminr-prod.