ai roadmap
A path for people who will ship LLM features. You learn enough of the stack to debug token limits, retrieval failures, and unsafe outputs — not only to paste prompts.
Stages run top to bottom; skills inside a stage can be taken in any order unless a line says otherwise. Locked nodes need their required predecessor first.
Python you can rely on and the network ideas transformers reuse.
How transformers actually work, and how to specify behaviour precisely.
Grounding models in your data, and when fine-tuning is the wrong answer.
Evals, red-teaming and the safety cases that let you deploy.