My project folders include the exercises behind later AI applications: TensorFlow notebooks, image-classification examples, text preprocessing, and sequence-model scripts.
These materials are useful parts of the history because they expose what happens underneath an API call. A model needs an input representation, a training objective, and a way to evaluate what it has learned.
Text preprocessing is part of the model
One retained NLP script fits a tokenizer, converts text to sequences, and defines an embedding-based Keras classifier.
The tokenizer creates the vocabulary-to-integer mapping. The embedding layer converts those integers into dense learned vectors. A later layer uses the vectors to make predictions.
These are separate operations. Giving a word an integer ID does not itself give the word a semantic representation. That distinction becomes useful again when dealing with LLM token counts and embedding APIs.
Images and sequences
The broader collection covers neural networks and convolutional-network exercises as well as text and time-series material. The common question is how the architecture should reflect the structure of the input.
Images contain spatial relationships. Sequences contain ordering. A model can lose useful information if the preparation step discards that structure.
Exercises also make evaluation more concrete: how validation data is separated, whether preprocessing learns from held-out examples, and which metric reflects the actual task.
Learning materials and later applications
The folders contain coursework and exam-preparation materials. They are evidence of a learning collection, rather than a certification or course-completion claim.
The technical connection is the move from manually selected features toward learned representations. Later LLM applications still face the same practical questions: what enters the model, what the output means, and how software checks it.
Any public examples should separate my own work from instructor materials and use small, attributed exercises with cleared outputs.