DisTec Challenge finalist: SADIE - A secure data environment powering an AI virtual coach

Defence is currently struggling with two diametrically opposed forces:   

• An exponential increase in the complexity of systems, information and data.  
• A reduction in the availability of specialists available to train others to use these systems effectively.

Therefore this is a need to augment our human instructors and learning systems, with a more intuitive and natural instructional system that can adapt to the learner to support their education.

The last 18 months have seen a surge in the democratisation and availability of generative AI tools like ChatGPT, Meta's Llama and Anthropic's Claude. The level of investment required to create such a model from scratch is prohibitive (ChatGPT owner OpenAI have spent more than $530m to date), however, a number of LLMs have been made available on commercial or open-source terms, and can be deployed on secure or air-gapped servers.  

While Defence may not be able to use the open internet versions of these LLMs due to the classification of reference data, if these LLMs are securely hosted and trained on the relevant data sets, they can be deployed on a number of defence use cases.  Mission Decisions has developed SADIE - a Secure Analytics & Data Intelligence Environment that can host the structured reference data, machine learning algorithms and Large Language Models that can be applied to a wide variety of Defence problems.  

One application we have developed with SADIE is a Virtual Coach application based on a number of machine learning techniques and leverages open-source LLMs. It has been trained on defence-specific data not available on the open internet, whilst also addressing the security, governance and assurance concerns that will make it a trusted and an enduring force multiplier for Defence Training and Force Development.  

Our data science team has developed a process that can prepare the relevant source data in order to train the LLM to act as a virtual coach or learning assistant, providing referenced answers to questions or acting as an agent to help the learner get confidence in a specific subject or explain complex concepts. In turn, this application can be used to support the existing instructional staff to deliver the more critical or nuanced courses, whilst improving the overall experience for students.