Most generative AI courses hand you a framework and a notebook. This one hands you the decision. By the end of this course you can take a vague AI feature request and return a defensible architecture: named layers, a named owner for each boundary, stated limits on latency and cost, and a release gate that is allowed to block a bad build. You can say which part of the stack a wrong answer came from, and you can say why a framework is in the design instead of a plain function.
The running example is a Document Copilot, an assistant that answers questions over private documents, cites the passages it used, calls one read only tool, refuses when the evidence is not there, and leaves a trace an engineer can read after an incident. Every section builds one more layer of that system and ends with independent practical work that produces a real artifact: a boundary map, a model selection note, an answer contract, an evidence plan, an evaluation table, a trace review. Each comes with self check criteria, so you can judge your own result without waiting for a grader. The final portfolio project assembles them into one architecture pack you can put in front of a hiring manager or a design review.
This is for Python developers and early career AI engineers who can already read a function, a JSON payload and an HTTP call, and who now have to make architecture calls in front of other people. It is for backend and data engineers asked to add retrieval, tools or an assistant to a product that already has users, and for technical leads who have to decide when a framework earns its place. It is not for complete beginners to Python, and it is not for anyone who wants a type along coding bootcamp.
