Delivery6 min read
The four-hour overlap rule
Distributed delivery does not fail on skill. It fails on the six hours a day where nobody can answer a question.
Document AI
Why Devanagari OCR is still hard in 2026
Conjunct consonants, inconsistent matras and scanned government forms — a field report from 340,000 documents a month.
- Published
- 28 May 2026
- min read
- 11 min read
- Written by
- The Attatva engineering team
All insights
The gap between the benchmark and the scan
General OCR reports excellent Devanagari numbers. Those numbers are measured on rendered text. The documents our clients actually process are photographs of forms printed on a dot-matrix printer in 2004, photocopied twice, then captured at an angle on a phone in a warehouse.
Conjuncts are not characters
क् + ष renders as क्ष — a single glyph whose visual form is not a concatenation of its parts. A model that segments by glyph has to learn hundreds of these as distinct units, and the long tail is genuinely rare in training corpora but not rare in legal names. Family names are where this bites hardest, and family names are exactly the field a customs authority will not tolerate an error in.
Matras float
Vowel marks sit above, below, before or after the consonant they modify, and the one that sits before is written first but pronounced second. On a skewed scan, an above-the-line matra drifts into the previous line's descenders. We spend more preprocessing effort on deskew and line separation than on the recognition model itself.
Two numeral systems, one document
Devanagari numerals and Latin numerals routinely appear in the same form, sometimes in the same field — a date in one, an amount in the other. Any pipeline that normalises numerals globally will silently corrupt values. We keep the numeral system as a field-level attribute and validate against the expected one.
What actually moved the numbers
In order of impact: aggressive geometric preprocessing, per-script fine-tuning rather than one multilingual model, field-level confidence instead of document-level, and a review queue that shows the operator the original crop. None of these are research contributions. All of them were worth more than a better base model.
Read next
Tell us what's breaking.
Send the problem, not a spec. We'll come back within one business day with an honest read on whether we're the right team — and who to talk to if we're not.
attatvasystems@gmail.com · Replies within 1 business day, Kathmandu time (UTC+05:45)