$10–11/hr · Mercor · Hourly, 20 hours a week
You annotate Gujarati PDF pages with structural components and transcribe text faithfully in Gujarati script for document AI training.
What you would do
- Identify and bound structural components on PDF pages: headings, paragraphs, tables, figures, formulas, captions
- Assign each component a type according to project taxonomy
- Establish reading order reflecting how the page is actually read, including multi-column layouts
- Transcribe content character-for-character in Gujarati script, faithfully capturing handwritten and printed text with diacritical precision
- Extract page metadata: language, document type, dimensions, and content flags
Who they want
- Native Gujarati speaker with full command of script, diacritics, and conjunct consonant forms
- Experience with document work: annotation, transcription, translation, localization, or proofreading
- Systematic approach that applies taxonomy consistently rather than improvising per document
- Comfort working with unfamiliar layouts: newspapers, exam papers, handwritten forms, multi-column designs
- Character-level accuracy priority: precision matters far more than speed
Main skills
What the interview asks about
1.Gujarati script accuracy
Non-native speakers miss diacritical marks consistently; AI trained on incorrect script produces gibberish when deployed, wasting months of collection effort.
For example: “You encounter a paragraph with ઘ (gha) appearing multiple times, some with and without specific diacritics. How do you transcribe each occurrence correctly, and what reference do you use to verify you haven't conflated similar glyphs?”
2.Multi-column reading order
Establishing correct reading order trains models to understand page flow; wrong order makes the transcription useless for document understanding.
For example: “A newspaper page has three columns, a banner headline spanning the top, and a sidebar advertisement. How do you establish reading order? Does the headline come first, then columns left-to-right, or are there other logical sequences?”
3.Handwriting vs. printed distinction
Models need to know which regions are handwritten versus printed to behave appropriately when deployed; conflating them degrades training.
For example: “A form has printed labels but some fields filled in handwriting. The handwriting is reasonably legible but unusually formed. Do you transcribe it as-is, flag it as handwritten, or mark it as unclear? Why?”
4.Component boundary judgment
Incorrectly bounded components misrepresent structure; a table that's marked as one component when it's actually three separate sections teaches wrong patterns.
For example: “A page contains three tables in a row, each with the same column headers but separated by blank space and captions. Are these three components or one? How do you decide?”
5.Page-level metadata
Metadata (language markers, handwriting presence, formula complexity) helps researchers understand what their model is seeing and whether it generalizes.
For example: “A page contains English and Gujarati text side-by-side, appears to be from a bilingual exam, and includes mathematical formulas. What metadata flags would you record, and which are most important?”
6.Legibility and unsuitable-page judgment
Flagging unsuitable pages early saves time and prevents training on pages too degraded to transcribe accurately.
For example: “You open a page that's clearly Gujarati but heavily water-damaged, with significant portions illegible. At what threshold do you spend ten minutes trying, versus flagging the page unsuitable and moving on?”
A task you may get
Annotate and transcribe a provided Gujarati PDF page with structural components, reading order, metadata, and handle handwritten sections, then review a colleague's annotation on a second page for accuracy.
How to prepare
- Review the component taxonomy and practice applying it to diverse document layouts so you're fluent in classification before time-pressured work
- Gather examples of Gujarati text in different contexts (formal, handwritten, printed, multi-column) to refresh your recognition of script variations and diacritics
- Prepare a template or process for flagging unsuitable pages so you can make the judgment consistently rather than ad-hoc
The facts
- Pay
- $10–11/hr
- Hours
- Hourly, 20 hours a week
- Where
- Remote
- Open to
- IND
- Field
- Miscellaneous
- Project name
- Neon
- Posted
- 9/2/2026
- Places left
- 28
We wrote this page from the public Mercor listing. It may be out of date, so read the full posting before you apply.