A Model in a Notebook Is Not a Product
There is a specific feeling every ML engineer knows. The notebook runs top to bottom, the validation score looks great, and for about an hour you feel like a genius. I have learned to treat that feeling as the starting gun, not the finish line.
FaceFind taught me this the hard way. It is a facial recognition system for photo retrieval, and in the notebook it was beautiful. Clean dataset, good lighting, faces looking politely at the camera. Then real photos arrived. Blurry group shots. Half a face behind someone's shoulder. Event photos where the lighting came from a projector. The model had never been to a real event, and it showed.
The model is maybe twenty percent of the work
Nobody tells you this in a course, so I will say it plainly. The trained model is a minority stake in the product. The rest is the pipeline that cleans incoming data without silently corrupting it, the API that does not fall over when ten requests arrive at once, the storage decisions, the fallback for when the model is not confident, and the interface that a non-technical person can use without reading anything.
NoteSync made this even clearer. On paper it is a lecture note-taking app: transcription, summarization with Gemini, and Urdu-English support. In practice the hard problems were never the AI parts. They were the messy human parts. Lecturers who switch language mid-sentence. Phone microphones in echoing classrooms. Deciding what a good summary even means when two students want different things from the same lecture. You do not solve those with a bigger model. You solve them by watching real people use the thing and being humble about what you find.
What I do differently now
Three habits stuck with me from these projects. First, I get an ugly end-to-end version working before I optimize anything. A bad pipeline that runs beats a great model that does not connect to anything. Second, I test with hostile data on purpose. The worst photo, the noisiest audio, the weirdest input a user could throw. If it only works on the demo data, it does not work. Third, I ship to a small group early. Five real users find more truth in a week than I can find alone in a month.
None of this is glamorous, and that is exactly the point. The gap between a notebook and a product is where most AI projects die. Learning to cross that gap, over and over, is the actual craft.
