It's just hundreds of variables of data and some calculated features - basically from current point in time and some configurable points in history + some aggregated history data, fed to ML.NET FastTree / FastForest. I think LSTM / some Python frameworks could have been better, but learning C# better was one of the main reasons so ML.NET algorithms it is. I get quite reasonable predictions, though I need to do some more to get better at predicting the end of an activity because of some limitations how I'm giving the data to avoid overfitting. GPT is just going to be last step of the process, being fed todays actions and we'll see if I can get it to blurt out anything interesting based on it.

I just checked my manual tags in the training database and I've been on this for 197 hours and 37 minutes since since 12th of October. No wonder I feel like I need a break..