What it was
My final-year project for the BSc in Electrical and Computer Engineering at Wolaita Sodo University. Cotton growers had no quick way to tell leaf diseases apart in the field, so I trained an image classifier on labelled cotton leaf photos using transfer learning on a pretrained ResNet in TensorFlow and Keras. The model served two front ends: a Flask web app where a user uploads a photo and gets the predicted disease class, and an Android app that runs a TensorFlow Lite export of the same model on the phone, so a diagnosis does not depend on a connection. The thesis, the training notebook and both apps are in the repos linked below.
What I learned
Transfer learning made a small dataset usable, but the gap between notebook results and field photos taught me more than the model did: lighting, background and phone cameras all shift the input. Exporting quantised and unquantised versions for the phone forced me to think about model size and precision long before I met the same tradeoffs in production. It was also the first time I shipped one model through two interfaces, a pattern I still use.