Scale AI Uses Reducto to Parse Government Documents for Its Donovan Agents in Air-Gapped Environments

Scale AI, known for supplying training data to frontier model developers such as OpenAI, Google and Meta and valued at $29 billion, also serves public sector and large enterprise customers through Donovan, a platform for deploying mission-tailored AI agents for government, defence and other high-stakes workflows, according to a December 2025 case study by Reducto. Donovan helps decision-makers organise and analyse large volumes of unstructured data, and its inputs are mostly government documents such as official orders, situation reports and intelligence briefs, which can contain handwriting, charts, labelled maps and low-quality scans. Before adopting Reducto, Scale's public sector team ran a homemade document-processing stack, building a one-off solution for each document type. "A customer will come in with a specific use case where our pipeline doesn't work as well as we would like. This led us to have to go in and manually support the new file," said Shourya Munjal, a software engineer on the team. As use cases multiplied, the team looked for a more efficient approach that would still protect confidentiality. The team now mainly uses Reducto's Parse endpoint to turn raw documents into clean text chunks that are embedded and reused across Donovan's retrieval-augmented generation chats and other mission applications; it started by integrating Reducto into its embedding pipeline. Reducto also produces figure summaries, so visual content is not lost. Deployment constraints were strict. The product had to run on premises in an air-gapped environment to protect sensitive government data, operate under tight GPU limits, integrate with approved cloud providers and work with the small set of models authorised for high-side environments. Reducto's on-premises option met these requirements. "We work with a lot of government customers who have very strict requirements around data security and privacy," said Kyra Huneycutt, product manager of Donovan.