Data Cleaning

What it handles

Pure scanned documents with no text layer — the hardest shape of archive data. Typewriter text mixed with handwriting, hand-drawn tables, formulas, drawings, stamps and signatures.

How it works

Three steps, one handoff. You never lose control of your documents — the conversion result belongs to you.

Case studies

A 1975 geological exploration report — 189 pages, 300 DPI scan, every digit verified by hand against the original.

Chemical analysis table · 120 cells

120/120 correct

Scan (input)

Structured output

%	15	699	59	14	1128	50	6	46	20
SiO₂	55.82	68.56	63.11	53.69	68.50	57.50	72.90	65.68	63.91
TiO₂	0.90	0.45	0.80	0.80	0.55	0.79	0.48	0.57	0.59
Al₂O₃	16.09	15.63	15.63	18.76	14.80	17.33	14.18	16.25	15.88
Fe₂O₃	2.25	2.53	2.12	3.27	2.96	3.78	1.65	2.38	3.22

Oxide column sums re-verified against geochemistry rules automatically. The hardest cell — Al₂O₃ 18.76 — was missed twice by the second engine and caught by the primary.

Handwritten formula page

converted · uncertain terms flagged

Scan (input)

Structured output

Δg = g·α²/2

Handwritten symbols converted to real math; anything the engine is unsure about is explicitly marked for human review instead of guessed.

Landscape geological profile

embedded + searchable caption

Scan (input)

Structured output

Drawings stay as images inside the structured document, with a generated caption so they become searchable — no attempt to redraw what was drawn by hand.

How it was verified

300 DPI crops audited cell by cell by hand, on a real 1975 report — not a synthetic benchmark.

The pipeline also caught and corrected 7 pre-existing errors in the client's own manually-edited reference data — including a lost entire row (п.п.п) and a misread city name.

Cost & speed