Viraasat: A Human-in-the-Loop Platform for Digitizing, Transcribing, Modernizing and Translating Historical Marathi Manuscripts

Marathi manuscripts such as royal orders, land grants, Modi-script records and medieval literature are deteriorating, and only a few specialists can read them. Scanning preserves them as images but does not make them searchable, and general-purpose OCR tools fail on faded ink, skewed pages, archaic orthography and Modi script. This paper presents the design and interim progress of Viraasat (Marathi for "heritage"), a full-stack web platform that takes a scan or photograph of a historical Marathi manuscript through a nine-stage workflow: ingestion, image preprocessing, text-line segmentation, AI-assisted transcription, human review, modernization of archaic Marathi, translation into Hindi and English, synchronized reading and export. Preprocessing combines deskewing, CLAHE contrast enhancement and edge-preserving filtering. Lines are located by signal processing on the Devanagari headline (shirorekha) using projection profiles, smoothing and peak and valley detection. Transcription is delegated to interchangeable providers (Google Gemini multimodal vision, Hugging Face Indic TrOCR and Tesseract), and doubtful words are bracketed with low confidence. A scholar reviews every line, every correction is stored in an audit ledger, and the original scan and raw OCR are never changed. The platform follows five archival design principles and is built as a three-tier system with a React front end, a Node.js back end, a Python AI service and MongoDB. Progress recorded up to October 2026 (about 40% of the project) comprises the design, preprocessing, line segmentation, OCR providers, back end, review interface and automated tests. Evaluation of recognition accuracy, validation of modernization and translation, and user testing remain pending, and no performance figures are claimed. The system is a research prototype, and its output requires scholarly verification..