Donovan

The Donovan platform, developed by the San Francisco-based data-labelling and applied-AI firm Scale AI, is a software system designed to integrate large language models into United States defence decision-making workflows. Introduced in 2023, the system functions as an “AI Agent” factory, allowing military and public sector users to customize, evaluate, and deploy specialized artificial intelligence tools for mission-critical tasks. Scale AI positions Donovan as a mechanism to fuse orders, situation reports, and intelligence data into natural-language answers and draft courses of action, aiming to reduce the time between intelligence gathering and operational planning.

Technically, Donovan operates on a retrieval-augmented generation (RAG) architecture, enabling it to query classified and unclassified data sources without retraining foundational models for every new query. The platform hosts the “Defense Llama” model, a customized version of Meta’s Llama 3.3 70B Instruct, which Scale AI states is fine-tuned for defence-specific contexts. The system ingests data from various connectors, including the Global Database of Events, Language, and Tone (GDELT) event data and Department of Defense policy documents, to build knowledge bases for its agents. These agents perform tasks such as real-time intelligence analysis, threat monitoring, and the extraction of location data from unstructured text. Scale AI highlights the platform’s ability to run on air-gapped and classified networks, citing authorization for DISA IL4 and FedRAMP High environments, alongside cloud-agnostic Kubernetes deployment options.

The primary operator of the Donovan platform is the United States government, specifically within the national security and public sector domains. Scale AI markets the system to US defence agencies, emphasizing its role in achieving “AI Overmatch” against adversaries. The company’s federal website details how Donovan assists in drafting operational reports and analyzing artillery capabilities, though specific contract values, deployment counts, or distinct unit assignments remain undisclosed in public materials. The system is not described as combat-tested in the traditional sense of fielded hardware; rather, it is a decision-support tool intended for use by analysts and commanders in secure environments.

Scale AI continues to iterate on the Donovan ecosystem, introducing features such as the “Agent Arsenal” for discovering pre-built intelligence tools and “Agent Factory” for no-code customization. The platform supports multiple model comparisons, allowing users to weigh metrics like factual accuracy and hallucination rates between different foundation models, such as Scale’s own Gemma 7B Instruct against larger Llama variants. By providing traceable reasoning steps and guardrails, Scale AI aims to address accountability concerns inherent in using generative AI for military planning. The system represents a shift toward software-defined command and control, leveraging commercial large language model capabilities to process vast amounts of intelligence data for US defence operations.