What engineers who have been through the courses say

The labelling quality exercise was the part I did not expect to spend so much time on. I came in thinking the model training was the hard bit. It turned out the dataset construction was where most of my mistakes were hiding. The code review caught specific issues in how I was writing the augmentation pipeline — things I would not have noticed on my own because the model still converged.

I came into this after deploying a RAG system at work that I could not honestly evaluate. The adversarial evaluation exercise was harder than I expected — building the set yourself is a different thing from reading about it. The harness I wrote during the course is now the same one we use in production. That was the most useful single deliverable I have ever produced in a course.

Ten to twelve hours a week was accurate — sometimes a bit more around the evaluation exercises. The weekly sessions were useful because the instructor was the same person reviewing my code, so the session answers were actually connected to what I had submitted. The edge deployment exercise was the most chaotic part, partly because of my own hardware setup, but the course materials helped me work through it.

The annotator agreement session was not something I had seen covered properly anywhere else. We had a disagreement in my team about how to handle ambiguous cases in our labelling, and the session gave me the framework to have that conversation properly. The public write-up was more work than I expected but it is now something I can share in any technical context without hesitation.

I work on a manufacturing line in Johor and the course scenarios were directly applicable — specifically the augmentation section on what hurts performance in an industrial setting where lighting changes and angle variation are consistent problems. I would have liked a bit more on multi-camera setups, but that is a specific thing, not something to read as a criticism of what the course covers.

The chunking strategy section was denser than I expected. I had to revisit several of the materials. The office hours were the right place to get that sorted — the instructor answered questions from the specific code I had written, not from a generic answer. The reranking exercise was where things clicked for me. The harness I built is genuinely useful at work now.