ICE Develops Internal Platform for Building AI Tools

The US Immigration and Customs Enforcement agency is developing a new internal platform designed to standardize how its teams build and test artificial intelligence applications. This system, named STELLA, aims to provide a secure environment for developers to experiment with advanced language models and other AI capabilities without leaving the agency's controlled infrastructure.

According to documents obtained by GN technics/ai (en-US) through a public records request, the agency describes the platform as a secure hub for developing, testing, and deploying new use cases. The tool is intended to give staff access to standardized software environments and security checks, ensuring that all AI experiments are conducted within established federal cloud boundaries.

Internal tools standardize AI development

STELLA, which stands for Secure Technology Environment for Large Language Agents, is designed to simplify the technical process of creating AI solutions. It provides developers with containerized environments, which are isolated software packages that make it easier to run experiments safely. The platform also includes built-in security scanning and logging tools to monitor how the AI systems behave during development.

The agency stated that the platform will handle authentication and authorization, meaning it manages who has access to which data and tools. By centralizing these functions, ICE hopes to reduce the risk of security vulnerabilities that often arise when individual teams build their own custom AI setups. The system is scheduled for an initial release this summer, although the exact date remains subject to technical finalization.

Deployment timeline remains uncertain

While the agency announced the platform in its AI inventory, the status has been marked as pre-deployment for several months. In January, the tool was already part of efforts to automate business functions, according to comments by ICE’s Chief Information Officer. However, when the inventory was updated in July, the project was still listed as being in the development stage.

ICE has not provided specific details on when the system will be fully operational. The agency responded to requests for clarification by stating that the architecture and models are still being finalized. This lack of a concrete timeline creates uncertainty about when the tool will actually be available for use by the agency's various teams.

Questions on accuracy and data

Significant questions remain about the quality and safety of the AI systems built on this platform. ICE did not disclose information regarding the accuracy of the models or their rates of producing incorrect information, a common issue known as hallucination. The agency also did not specify which external companies were involved in the development or what types of immigration data were used to train the tools.

STELLA is one of more than forty AI projects currently in use or development at ICE. These include tools for analyzing mobile devices and processing tips. However, the agency has recently retired several other AI applications, such as a resume screening tool and a system for evaluating community residence. This suggests a cautious approach, where projects that do not meet performance standards are removed from service.