I'm drawn to the parts of machine learning that people skip. Anyone can call model.fit(). I'd rather understand what it's doing — so I wrote my own deep learning framework, backpropagation and all, before I let myself use TensorFlow.

That habit shows up in how I build. My sentiment project didn't start with a dataset; it started with a scraper, because the dataset didn't exist. When labeling by hand proved impractical, I used an LLM as a weak supervisor and checkpointed the whole run so it could survive an API failure at 3am.