> For the complete documentation index, see [llms.txt](https://hasithz.gitbook.io/hasith-abayakoon/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://hasithz.gitbook.io/hasith-abayakoon/selected-work/projects/drdok.md).

# DrDok

Document digitisation for scanned forms, images, and PDFs at BooleanLabs

**Research Engineer / Technology Consultant, BooleanLabs, May 2023 – May 2024.** Applied AI for document processing.

<figure><img src="https://1402428123-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FnSR0OqsIyopxLb9jZNtc%2Fuploads%2FI26HG6KVzpL6ed5pBhPB%2Flogo-logindrdok.png?alt=media&amp;token=2990f966-9479-45ac-b9bc-cf0637100ad2" alt="DrDok logo" width="320"><figcaption><p>DrDok, the document digitisation product from BooleanLabs.</p></figcaption></figure>

## Problem

Organisations hold information in scanned documents, photographed pages, and PDF files, in both structured forms with fixed fields and unstructured documents with free layouts. Extracting that information by hand is slow and error-prone, and layouts vary in size and structure from one document to the next. DrDok brings these inputs into a single digitisation workflow so the content can be used digitally.

## My role

I worked on the recognition models behind the product: developing and fine-tuning convolutional and graph neural networks and evaluating vision transformers for image recognition on the product's documents.

## Constraints

* **Varied inputs.** Images, scans, and PDFs of different sizes, resolutions, and quality, from structured forms to free-form documents.
* **Layout carries meaning.** Where a field sits on the page matters, so recognition could not treat a document as a flat sequence of characters.

## Decisions

* **Compare model families rather than commit early.** Convolutional networks, graph neural networks, and vision transformers were each developed or evaluated for image recognition on the product's documents, so the choice rested on measured performance.
* **Fine-tune rather than train from scratch.** Existing networks were fine-tuned to the document domain.

## Evidence

* Product: [DrDok at BooleanLabs](https://booleanlabs.biz/)

## Outcome

The models I worked on fed the product's digitisation workflow. The product's current status and any measured results are not published here.

[Back to selected work](/hasith-abayakoon/selected-work/projects.md)