Alright, thanks for sharing.
What I think you will need to do is provide the model with a condensed version with the relevant data only. And this you can then much easier analyze in a timely fashion via a regular script.
In other words I don’t see any additional value a AI model can add to the analysis. Unless you feed large amounts of data into a model which in turn will look for patterns in the data that you cannot see.
On the other hand you can definitely use an LLM to help you dissect the logs and create the inputs for an analysis.
Lastly, what is sometimes done , is taking a screenshot every other frame and feeding it into a vision model for analysis. But again, this will be slow and the model needs to know what to look for and would require annotations on the fly and in a fast paced game you will miss relevant information.
I would stick to the classical log analysis and feed the data into a regular script to provide you with real time feedback about regular KPIs.
Then you can save all the data and use it later to show it to an AI looking for new insights.
Maybe? That’s my thought process so far.