Tramscribe

Who is Tramscribe

Tramscribe is a recent startup aiming to save time for NHS clinicians by automating admin tasks. Clinicians within the NHS often have to take many notes and spend time doing administrative tasks that distance them from patients.
Tramscribe speeds up this process by using AI to transcribe medical voice notes and then turns the raw notes into well-formatted templates. However, ensuring data privacy when sending patient data to outside AI solutions is a real challenge. Tramscribe was looking for a solution to leverage the latest AI models with Large Language Models while respecting patient privacy.

Main findings

• Tramscribe developed a state-of-the-art medical voice note transcription solution using LLMs.
• Tramscribe was able to successfully deploy those AI workloads inside BlindBox to ensure data privacy for patients.
• Leveraging Cloud infrastructure on Azure with Confidential Containers reduced time to production while ensuring little exposure of patient data.
• Tramscribe can now provide a precious time-saving solution for clinicians with a fully managed app and privacy guarantees.

Challenges

Tramscribe wanted to provide a solution that would be as easy as possible to use by clinicians. Therefore, an on-prem solution would not meet those requirements. A fully SaaS solution in the Cloud would be the best choice for onboarding and infrastructure management.
In addition, because Tramscribe wants to provide AI-based transcription and summarization, getting the latest hardware and software from a Cloud provider is the natural way to go.
However, proposing a SaaS solution implies exposing data to Tramscribe Cloud platform. Then guaranteeing that patient data sent to Tramscribe would not be exposed is a real challenge. Indeed, Tramscribe would centralize patient data and can become a point of failure regarding this sensitive information.
Consequently, reconciling privacy and security, along with the imperative for a swift onboarding mechanism that accommodates an advanced AI infrastructure, presented a substantial technical challenge.

Our solution

By using BlindBox, it was possible to deploy AI models on a Public Cloud while guaranteeing data privacy by leveraging Confidential Containers.
We deployed two open source models, Whisper from OpenAI, to perform speech-to-text, then used a ChatGPT-like model, Pythia from TogetherComputer, to conduct medical notes structuration. Both models were packaged inside a Docker container which was then hardened using BlindBox and deployed on Azure Confidential Containers.
Subsequently, we utilized the BlindBox client SDK to query these AI models, thereby maintaining end-to-end data protection. This approach ensured that patients' data remained secure, as key management and control verification were handled on the client side, preventing any unintended exposure of sensitive information.