Some of the most important work in public service happens in places most people never see. It happens in back offices, in workflows, in forms, in the small systems that shape whether people can actually access the institutions meant to serve them.
At the AI + Elections Hackathon, our team built a tool focused on one of those systems. We created an AI-powered accessibility engine that takes scanned or handwritten government documents and transforms them into fully accessible, ADA-compliant digital outputs. It was built for election offices and government agencies that carry a clear responsibility: public documents should be accessible to the public.
That sounds obvious. In practice, it often is not.
The Accessibility Gap
n April 2024, the Department of Justice finalized new rules requiring state and local governments to meet WCAG 2.1 AA accessibility standards. The original compliance deadline was April 2026, though the DOJ has since extended it, giving offices more time to bring their workflows into line. Many are still operating with systems built long before digital accessibility became a legal and civic priority. We heard this directly from election representatives during the hackathon.
A paper form comes in. It gets scanned, it gets saved, and that is often where the process ends. No OCR, no tagging, no accessibility review, no remediation. The document exists online, but for someone using a screen reader, it may as well not exist at all. A scanned image is not access, and a public document that cannot be read is not truly public. That was the problem we set out to solve.
What We Built
Our tool accepts scanned PDFs and handwritten forms and converts them into multiple accessible, machine-readable outputs without requiring manual retyping or remediation from the user.
It produces:
A PDF/UA compliant tagged PDF
A clean, styled DOCX file
Structured JSON ready for database integration or public data portals
The goal was compliance, usability, trust, and removing barriers for public servants and the communities they serve.
How It Works
A scanned PDF goes in. Document AI extracts the text and detects layout, then every content block runs through a confidence check. Anything below 95% confidence moves through a Gemini second-pass recovery, and anything still uncertain is flagged for human review. From there, the content is enriched with the accessibility structures that matter: proper heading levels, table structure, alt text, language metadata, and semantic tagging for screen readers.
The output is rendered into a tagged PDF/UA and a clean DOCX. Every file is validated against PDF/UA standards using veraPDF, and if validation fails, an AI fix loop catches and corrects the issue automatically. Only after everything passes internal validation does a human make the final decision before publication or download.
That final step matters. AI does the heavy lifting, but humans hold the responsibility. In election work, trust cannot be automated.
What We Learned
We entered the hackathon expecting to work on something broad. Instead, we found ourselves focused on a very specific workflow that carried enormous consequences. That shift happened on day one. Sitting with election representatives and hearing how handwritten forms were handled changed everything. It reminded us that innovation does not always begin with invention. Sometimes it begins with listening closely enough to notice where people have learned to work around broken systems.
The challenge was not discovering a futuristic problem. It was recognizing an existing one that had been normalized, and that realization shaped the project.
The Better Questions
One of the most valuable parts of the experience came from the judges. Their strongest questions were not about the technical architecture. They asked about adoption, about accountability, about legal responsibility. Who verifies accuracy? What happens when confidence scores are wrong? How does an election office actually implement this? Who owns the final decision?
Those were the right questions, because public-sector technology is never just about whether something can work. It is about whether people can trust how it works. The project became stronger because of that. It pushed us beyond building a tool and into thinking seriously about governance.
Why This Matters
This project matters because election offices and public agencies are already doing an enormous amount with limited time, staffing, and resources. Accessibility is often one more critical responsibility added to an already full system. What we saw during the hackathon was not a lack of care. It was a workflow built for efficiency in a fast-moving environment where people are trying to serve the public well with the tools they have. A form comes in, it gets scanned, it gets uploaded, and the work keeps moving. The challenge is that accessibility can easily get lost in that process, not because it is unimportant, but because the system was not originally designed with it at the center.
That is where this project matters. The goal was never to create more work for already stretched offices. It was to support the people doing that work by making accessibility easier, faster, and more realistic to achieve. Public documents should be truly public. That means voters using screen readers, community members accessing local government information, and anyone who deserves independent access to the same information as everyone else.
This is about strengthening trust. When institutions can make information accessible, they are reinforcing transparency, inclusion, and public confidence. Our hope was to build something practical that helps offices meet that responsibility with less burden and more confidence. Sometimes improving democracy looks like major policy decisions. Sometimes it looks like making sure a single document can finally be read by everyone it was meant for.
