◇ Could this help me?
Yes. Reproducing a paper is one of the few portfolio projects that actually demonstrates something, because it requires reading a technical claim precisely, translating it into working code, and then proving the result matches, which is much harder to fake than a tutorial clone. Using an assistant to decompose the paper into tasks and later to diff the implementation against the source is a sound division of labor, as long as the comparison step is treated as a draft to verify rather than a verdict, since a model summarizing whether code matches a paper is exactly the situation where it will confidently miss a subtle difference. General practice.
I want to reproduce a research paper from scratch as a portfolio project. Read the paper I give you and break it into implementation tasks I can work through one at a time, ordered so that each one produces something runnable. Do not write the whole implementation. As I finish each task, review just that piece against the relevant section of the paper. When everything is implemented, do a full comparison against the paper and list anything missing, different, or ambiguous, and separate the things you are confident about from the things I need to verify myself. Finally, suggest three extensions, such as another dataset, another model to compare, or a hyperparameter study, and tell me which one adds the most to the project.