The problem
Banning AI pushes researchers to use it anyway, on personal devices and public services. Allowing it means sensitive data may be sent to a third party and used in ways you can't control.
The TrustBoxx approach
A chat assistant available only inside the TRE, connected to models in AWS Bedrock in your own environment. Researchers get AI help; the data stays behind the same walls as everything else.
How it works
1
Inside the workspace
Researchers use a familiar chat interface (LibreChat) from their TRE session. It isn't reachable from outside.
2
Through a gateway you control
Requests go through an internal model gateway (LiteLLM), which decides which models are available and records usage.
3
To models in AWS Bedrock
Models run through AWS Bedrock. AWS does not use your prompts or outputs to train models.
What researchers use it for
- Writing and debugging R and Python code
- Explaining statistical methods and output
- Drafting methods sections and summaries of results
- Working through documentation for packages available in the TRE
Controls that still apply
- Same boundary: anything the assistant produces stays inside the TRE. It can only leave through the safe outputs workflow.
- Your choice of models: only models you approve are enabled.
- Region: model availability depends on the AWS region of your environment. We confirm which models are available in your region before setup.
- Usage cost: AI usage is part of your cloud usage and appears in your monthly cost report.
Private AI is an assistant for researchers, not a way around the controls. It sees what the researcher shows it, inside the TRE, and nothing leaves without approval.