PDF Annotation & Transcription Experts – Malayalam

Mercor is assembling a team to create high‑quality training data for document‑understanding models that struggle with less‑common scripts. The position focuses on Malayalam, one of several Indic languages being targeted, and involves turning publicly available PDF pages into fully annotated, accurately transcribed resources. Each workday centers on locating a Malayalam PDF that contains at least one multimodal element such as an image, table, diagram, or handwritten section, then documenting its source. You will demarcate every meaningful component on the page—titles, headings, paragraphs, lists, tables, figures, captions, formulas, questions, answer fields—and assign a type and reading‑order index. Relationships between elements are recorded through parent identifiers, and the entire visible text, including any handwriting, is reproduced verbatim in Malayalam script, with illegible areas flagged. Page‑level metadata like language, document category, dimensions, and presence of tables or formulas is also captured. After completing a task, a second Malayalam expert reviews the work end‑to‑end, and seasoned annotators may perform these reviews themselves. The role requires native‑level Malayalam proficiency, complete familiarity with the script’s diacritics and conjunct forms, and prior experience handling documents—whether through annotation, transcription, translation, localization, proofreading, journalism, or regional‑language data review. Precision is essential; even a single incorrect diacritic is considered a defect, and a systematic application of a predefined taxonomy across many pages is expected. Comfort with unconventional layouts such as multi‑column newspapers, exam papers, and handwritten forms is also mandatory. Additional background in regional‑language AI data work, MTPE, subtitling, OCR correction, or Malayalam typesetting is advantageous but not required. The project emphasizes real‑world Malayalam documents rather than synthetic examples, and success is measured primarily by annotation accuracy, with handling time tracked as a secondary metric.