The state of Indiana has taken a cautious approach to generative AI following its February 2024 AI policy. SOS, specifically, has prioritized AI-enablement within its divisions, recently launching two AI-powered chatbots to support the user experience.
The office’s latest innovation is known as “Captain Record.” This tool was built to help SOS obtain insights from more than 100 years of documents, making unstructured data searchable for staff, said Robert Fulk, CIO of SOS. A lot of this data was previously “locked up” in PDF formats, he said. The new tool makes tens of millions of records searchable with AI, from business filings to auto dealer records to historical licensing information. These could include PDFs and handwritten documents from decades past. SOS staff can communicate in plain language to ask the model a question or find information.
The state is at 20 million pages in this repository and growing, Fulk explained. Some old files are still being added; and notably, as SOS files new documents, those will need to be added, too. Currently, the database is approaching 10 terabytes and counting.
“This wouldn’t have been possible without the latest AI tools and technology and the ability to scale,” he said.
The data in these records is now publicly accessible by constituent request per legislative mandate, and things that used to take weeks or months to find can now be found in minutes.
The new solution, launched March 25, was built on Google Cloud Vertex AI using the Gemini 2.0 family of models. The project is powered by mavQ’s automation platform.
Google’s AI engine ran nonstop for weeks to ingest this volume of data, and as it was imported, the model was trained on it.
It was a “herculean effort” to train the model, with documents completed in different ways — handwritten or electronically — and saved in different formats, Fulk said. He noted that the AI can not only understand that file differentiation, but also categorize documents.
“The AI can actually read and understand, so you can interact with the data,” he added. This piece will be developed further in Phase 2.
The goal for Phase 2, which Fulk hopes to see live in the next six months, is to redact personally identifiable information as needed, to make the tool accessible for public use. Phase 2 will also enable the model to interpret and understand these documents to turn unstructured data into structured data, from which new insights can be garnered, like demographic trends over time. This will give SOS access to information about the state and cities within it that never has existed before, Fulk said. The question going forward, as he put it, is how all this data can now be used.
As Fulk said, this government data belongs to constituents: “We’re here to serve the people, so [Captain Record]’s able to serve them and provide transparency and access to things as we move forward.”
The tool is expected to act as a model on providing data access and transparency to constituents. Within the state, he said, the endeavor acts as a proof of concept for other departments looking to improve their archives access.
If other departments take on their own version of this project, Fulk said he recommends developing good partnerships and having the source at the table for planning discussions. Specifically, he said that having an AI partner made the project successful because AI is constantly changing.
Google was selected in part for its translation capabilities; looking ahead, the state is offering some documents in multiple languages to better serve constituents, so this tool may also enable state staff to interact with multilingual records.
SOS is already looking to other AI implementations, including agentic AI, to better serve the public.
