Each project will have an overall objective, like run a compute vision model on a drone and show how it can track large mammals in real-time. Additionally, projects will have a small number of research questions that we want to answer within the field of MLOps. For example:

• How does our approach to training the model impact energy and thermal profiles when deployed to the hardware?
• Can our MLOps pipeline test model performance in terms of energy and thermal profiles?
• What tooling can we use to evaluate the performance of the model in such a way that it replicates the real-life deployment environment? i.e. if we’re running this model on a fast-moving drone, does our model evaluation step expose the model to inputs that would replicate this?

The research questions that we put into a project should be heavily influenced by our understanding of a couple of things:

1. Unsolved problems that we are aware of from our understanding of the industry and current literature.
2. Challenges presented to us by current and prospective customers, as well as partners

For example, the power and thermals question comes from a specific conversation I had recently with AI researchers in the public sector.

By framing projects around research questions we can readily align our projects with the format that HRMC uses when assessing R&D tax relief claims: here, the expectation is that R&D projects seek to resolve a set of uncertainties that are not readily resolvable by a “suitably competent professional”. It follows as well that in our ways of working, we want to plan projects to fit in with R&D tax claims in advance.