Well, first of all, Log loss is the metric stated by Zindi to be used here. Secondly, even though you chose to use "Accuracy" as your metric, still i could see an overffitted score of 99.98% of yours. I think the way forward is for you to set an "EarlyStopping" and at the same time use another metric entirely to monitor your chosen metric (say "Accuracy") while training your model.

Hope this helps.

One important thing to check is the order of the classes during submission. The classes has to be in the same order as found in the sample submission. I remembered falling into this trap in my very early submissions.