fennec.sh

Yanis Ghidouche · Fractional CTO

MenuClose

AI & Innovation

PDF Accessible.

Market solutions promised automation and stalled at 30% compliance. We reached 85% by deciding, document by document, what needed algorithms, symbolic AI or LLM Vision.

85%

compliance with the strictest accessibility standards

What was at stake.

Making PDFs accessible for hundreds of companies means dealing with documents that look nothing alike. Structure, layout and content change. A method that works on one file can fail on the next.

Promises of total automation irritate me when the result is still unusable. We needed to understand the accessibility standards and the documents well enough to choose the right treatment.

What needed to change

Market solutions promised 100% automation but delivered 30% quality. Document formats are too varied for a single approach.

We needed to adapt processing to each type of document.

What changed.

  1. Accessibility

    BeforeInadequate off-the-shelf solutions

    After85% compliance

  2. Companies served

    AfterNewHundreds

  3. AI architecture

    BeforeSingle approach

    AfterHybrid 3-tier

The choices I had to argue for.

  1. 1

    Decision 1

    I argued for a three-part hybrid approach. For structured content, deterministic algorithms do the job: their behaviour is predictable and their results can be checked.

  2. 2

    Decision 2

    Symbolic AI recognises patterns in semi-structured layouts. LLM Vision handles complex layouts that need visual understanding.

  3. 3

    Decision 3

    I want to justify each layer by the problem it solves. Adding a model where an algorithm is enough doesn’t magically improve a product.

The takeaway.

Knowing when to leave AI out is part of my job. On PDF Accessible, combining techniques is what let us handle that variety of documents. Judge the choice by the result, not the model in the announcement.

To connect compliance, product and technical choices, see my work on digital accessibility.