I wanted some ideas for my final year ai project

For something that could touch on new science as well as having commercial potential, you could create a rolling context non-generative audio AI model, that would replace small embedded device models in home assistant devices, which currently are using the phonic shape of wake words, technology a decade old. This instead would offer intelligent always-listening trained AI frontend across new voices and world languages that can understand when it is being addressed within natural language, without needing a large power envelope for processing any detected noise or even constant chatter, and able to intelligently forward and initiate an AI conversation session on its own from the correct start of buffering.