Edge Scriptorium - Accessible, Privacy-First Capture for Ancient Texts and Complex Scholarly Documents

Project Proposal for The Prompt Challenge

I am a disabled independent researcher developing a local humanities knowledge architecture called the ∑-engine.

Handling books, repeatedly rescanning pages, and correcting poor captures can be physically exhausting or sometimes impossible. Existing book scanners can capture an image, but they do not reliably determine whether a page is stable, properly illuminated, unobstructed, correctly framed, or ready for further scholarly processing.

My proposed project, Edge Scriptorium, adds a privacy-first edge-AI capture and document-structuring layer to overhead book-scanning systems. It is designed for ancient texts, books, facsimiles, scholarly editions, annotated pages, and other complex page layouts.

The Missing Layer Before OCR

Edge Scriptorium is not intended to solve OCR itself.

Instead, it addresses the difficult capture-and-structure stage that comes first:

• detecting a page, open book, facsimile, or document;
• identifying page boundaries and relevant visual regions;
• waiting until hands and motion have cleared;
• detecting and, where possible, digitally removing visible thumbs or fingers when they are needed to hold a book open, while flagging any obscured text for recapture;
• assessing stability, sharpness, lighting, shadows, and glare;
• selecting or approving the best available capture;
• correcting perspective and page geometry;
• enhancing the image where appropriate;
• segmenting text lines, illustrations, annotations, or sign regions;
• exporting clean image regions with quality and provenance metadata.

This creates more reliable input for later OCR, transcription, annotation, archival work, or scholarly analysis.

Why Edge AI?

The system is designed to operate locally on the Dell and Axelera Metis platform, with host-side processing used where appropriate.

Book images, annotations, unpublished research, and copyrighted study materials should not need to be uploaded to a cloud service merely to determine whether a page has been captured correctly.

Local processing provides:

• immediate feedback while scanning;
• privacy for books and research materials;
• offline operation;
• low-latency capture decisions;
• predictable control over stored images and metadata;
• a reusable foundation for future scholarly tools.

The output would consist of clean page images, segmented regions, quality measurements, and provenance metadata that could later be used by OCR systems, digital-humanities tools, archival workflows, or a local research knowledge engine.

One-Month Proof of Concept

The initial prototype would use:

• my existing NETUM T101 A3 overhead book scanner as the primary capture device;
• Voyager Wingman to generate, test, and iterate the pipeline;
• a cascaded Metis computer-vision pipeline;
• local host-side logic for capture control, metadata, and export where appropriate.

The NETUM scanner is being repurposed as a practical starting point rather than treated as a requirement.

Edge Scriptorium is intended to remain camera-source agnostic. The same capture-and-analysis pipeline could also work with a low-cost homemade two-camera book-scanning jig, with one camera directed at each page of an open book.

This would make the system reproducible for researchers, makers, small archives, and disabled users who do not own a commercial book scanner.

The demonstrable MVP would:

1. detect a page or open-book spread and identify its boundaries;
2. recognise whether a hand or significant motion is still present;
3. evaluate image stability, focus, lighting, shadows, and glare;
4. select or approve the best available frame;
5. correct perspective and enhance the captured page;
6. segment text lines or other visual regions;
7. export the processed images and metadata locally;
8. display the capture state in a simple local interface or event log;
9. detect and, where possible, digitally remove visible thumbs or fingers used to hold a book open, while flagging any obscured text or regions that require recapture.

Example events could include:

page_detected
hand_present
image_stable
focus_acceptable
glare_detected
best_frame_selected
perspective_corrected
regions_exported

Public Demonstration and Privacy

The public demonstration would use:

• self-created test pages;
• public-domain or openly licensed facsimiles;
• printed scholarly-style layouts;
• synthetic examples containing columns, notes, illustrations, transliteration, or unfamiliar scripts.

It would not expose private books, personal annotations, copyrighted course material, unpublished research, or sensitive documents.

Original manuscripts or archaeological objects are not required for the proof of concept. The pipeline can be demonstrated using facsimiles and complex printed material while remaining applicable to ancient-text research.

Why I Am Proposing This

I bring the lived use case, scholarly workflow design, research architecture, and clear privacy boundaries.

As a disabled researcher, I understand how a seemingly small task - turning a page, checking focus, rescanning, correcting perspective, or reorganising files - can become a major physical barrier when repeated across an entire book or research corpus.

Edge Scriptorium is intended to reduce that repeated handling and turn an ordinary scanner or camera rig into a more accessible, intelligent research instrument.

I have previously been in contact with Axelera AI about an edge-AI navigation and accident-prevention concept for mopeds. Edge Scriptorium is a different use case, but it reflects the same practical approach:

Local AI designed around real accessibility, privacy, and safety needs.

Voyager Wingman and Metis would make it possible to turn this architecture into a working edge-AI pipeline within the challenge period.

Longer-Term Direction

After the challenge MVP, Edge Scriptorium could grow into a broader local capture platform for:

• ancient-language editions and facsimiles;
• dictionaries and grammars;
• annotated research books;
• archival and historical documents;
• complex multi-column layouts;
• cuneiform, hieroglyphic, or other sign-rich material;
• commercial scanners, ordinary cameras, and affordable dual-camera DIY capture rigs;
• integration with the ∑-engine and other local scholarly systems;
• accessible hands-free capture controls using companion edge devices.

The first goal remains deliberately focused:

Create a working, privacy-first edge-AI layer that turns difficult page captures into clean, structured, provenance-rich research objects.