ScriptAIX

ScriptAIX is an AI-assisted transcription system for historical documents. It turns scans, photographs and PDF files containing difficult handwriting into readable, searchable text.

The software was developed primarily for genealogy and archival research. Instead of using one generic instruction for every document, ScriptAIX combines modern multimodal AI models with specialised prompts for particular periods, regions, languages, document types and handwriting traditions.

It can process anything from a single letter to a large folder of archival scans. The results are saved as Markdown and PDF files and can be reviewed alongside the original images in an interactive viewer.

ScriptAIX is currently in closed beta. It is being tested with real documents from genealogists and archive researchers, and its prompt library is continuously expanded using difficult and unusual examples.

ScriptAIX began with my own genealogical research.

Archives in Belgium, the Netherlands, Germany and France contain enormous quantities of handwritten material: civil registers, notarial deeds, estate declarations, church books, marriage registers, correspondence, population records and administrative documents.

Reading these sources requires palaeographic experience. Even after learning a particular handwriting style, the next archive, clerk or century may look completely different. Historical documents also contain much more than continuous text. They include marginal notes, corrections, deletions, insertions, tables, stamps, pre-printed forms, abbreviations and text written in several directions.

Standard OCR software is not designed for this kind of material. General-purpose AI transcription can help, but the results improve considerably when the model receives precise instructions about the document type, historical context, language and expected structure.

ScriptAIX turns those instructions into a repeatable workflow.

ScriptAIX can produce two different forms of transcription.

A diplomatic transcription stays as close as possible to the original document. It preserves the original spelling, abbreviations, punctuation, line structure, corrections and uncertain readings. This version is intended for researchers who need to verify exactly what appears on the page.

A normalised transcription is designed for easier reading and genealogical use. Abbreviations may be expanded, historical spelling can be made more accessible and the text can be presented in a clearer modern structure without changing its meaning.

Some presets produce both versions automatically. This makes it possible to compare the faithful transcription with a more readable interpretation.

The quality of a transcription depends heavily on the instructions given to the model. ScriptAIX therefore includes a growing library of specialised presets.

These cover, among other things:

• nineteenth-century Dutch administrative handwriting
• memories van successie and other inheritance records
• seventeenth-century Dutch secretarial handwriting
• marriage and banns registers from Zeeuws-Vlaanderen
• early modern Dutch and Flemish correspondence
• Belgian notarial documents from the seventeenth and eighteenth centuries
• Flemish parish registers written in ecclesiastical Latin
• French handwritten letters and administrative records
• German Kurrent and Sütterlin handwriting
• German, Dutch and Latin church books
• pre-modern forms, tables and mixed printed and handwritten documents
• a general preset for documents that do not fit an existing category

The presets do more than identify individual words. They tell the model how to handle page layout, printed text, handwriting, marginal additions, deletions, corrections, names, dates, monetary amounts, tables and common historical abbreviations.

Several presets have been substantially improved during the beta tests, particularly for complex tables, corrections made by the original writer and documents containing multiple text layers.

New presets can be added without changing the Python program itself.

ScriptAIX is not tied to a single AI provider.

The software can work with models from Anthropic, OpenAI and other compatible providers. Different models have different strengths. One may perform particularly well on seventeenth-century Dutch handwriting, while another may be better at tables, faded scans or German Kurrent.

The model can therefore be selected for each run or for each prompt preset. This also makes it possible to compare results and to take account of differences in speed, price and transcription quality.

ScriptAIX accepts common image formats as well as multi-page TIFF and PDF documents. It corrects image orientation, prepares the pages at a suitable resolution and sends them individually to the selected AI model together with the appropriate expert prompt.

Pages are processed in sequence and can also be handled in parallel. Large documents are loaded page by page, so they do not need to fit into memory all at once.

Failed pages can be retried automatically. Existing results are skipped when a run is resumed, and the progress of larger jobs is stored on disk.

For every page, ScriptAIX records metadata such as:

• the original source file
• the processing date
• the selected model
• the prompt preset
• the transcription type
• token usage
• estimated or actual processing costs

At the end of a run, ScriptAIX produces a cost summary. A dry-run mode can estimate the likely cost before the documents are sent to an AI provider.

Every local run can generate an interactive HTML viewer.

The original scan is displayed next to the transcription, making it easy to compare the result with the document. The viewer supports zooming, panning and searching across the processed pages.

Uncertain readings, illegible passages, marginal notes and editorial remarks can be highlighted visually. Metadata shows which model and preset were used for each page.

When the local server mode is enabled, the transcription can be corrected directly in the browser. Changes are written back to the Markdown file, while backups make it possible to restore an earlier version.

The transcription data is also stored in a separate structured JSON file so that it can be reused by other programs.

ScriptAIX now also includes an email agent.

Approved beta testers can send one or more document images or PDF files as email attachments. The name of the desired prompt preset is entered in the subject line. If no preset is specified, the standard preset is used.

The email agent processes the attachments automatically and returns the transcription by email. Results are currently supplied as both Markdown and PDF files.

An email with the subject help returns the current instructions and the list of available presets.

The email workflow is particularly useful for people who want to test ScriptAIX without installing Python, configuring API keys or using the command line.

Because email programs do not always preserve attachment order in the same way, especially with some Outlook configurations, page sequence remains an important area of testing and improvement.

ScriptAIX contains several features for processing larger archival collections.

A dry run lists the pages that still need to be processed and estimates the likely API cost before anything is submitted.

Parallel processing can speed up interactive runs. Prompt caching reduces repeated input costs where supported by the selected provider.

For compatible Anthropic models, batch processing can send larger collections through the Message Batches API at a reduced price. Batch status is saved locally, so processing can be interrupted and resumed without submitting the same pages twice.

The software also keeps source files, transcriptions and metadata clearly separated, making it easier to archive, verify and reuse the results.

ScriptAIX is intended to become open-source software and will be free to install and use locally.

Running the program locally does, however, require access to an AI provider. The provider charges for the tokens processed by its models. ScriptAIX records these costs transparently and includes tools for estimating them in advance.

The hosted email version is intended to remain simple and accessible. Users will only be charged for the AI tokens used for their own transcriptions, rather than paying a general software subscription.

ScriptAIX does not replace a palaeographer or archival researcher. It produces a first transcription that still needs to be checked against the original document.

Names, place names, dates, numbers and monetary amounts are especially vulnerable to misreading. Damage, ink bleed-through, poor photography and corrections by the original writer can also confuse even the strongest models.

For that reason, ScriptAIX is instructed not to hide uncertainty. Doubtful readings are marked, illegible passages remain visible as such and unusual features may be explained in a short remarks section.

The objective is not to create an apparently perfect text. The objective is to produce a useful, transparent draft that can be verified much faster than a transcription made entirely from scratch.

ScriptAIX has entered a closed beta phase.

A small group of genealogists and archive researchers is currently testing the software with real historical documents. Their examples and feedback are being used to improve page ordering, document analysis, transcription conventions, the user interface, the email workflow and the specialised prompt presets.

I am particularly interested in documents that are difficult to decipher or that produce inaccurate results. Unusual handwriting, damaged pages, complex tables, marginal notes and corrected text are especially valuable for improving the system.

ScriptAIX is still developing rapidly, but it is already capable of turning many difficult archival documents into a practical and verifiable first transcription.

If you are curious about the project, email me.