LLM Dojo is a free, in-browser course for engineers who need to understand language models well enough to ship them and to answer AI engineering interview questions. Aditya Jha, a full stack developer, built it as the hands-on reference he wanted when he started, with lessons you can manipulate, covering tokenization, attention, RAG, agents, evals, and inference. Seven modules and 41 lessons are open, about 15 hours from how a network learns through routing, fallbacks, quantization, caching, and cost per request. Lessons include views of the same text under gpt2, cl100k, and o200k tokenizers, MiniLM embeddings, and individual attention heads, along with sampling, context windows, and the KV cache. A playground trains a small network entirely on your device, on XOR, circles, clusters, and spirals, so you can watch gradient descent move the boundary. Interview practice sits on the same site: 492 questions, 78 junior, 224 mid, and 190 senior, tagged to write-ups from 58 companies including OpenAI, Anthropic, Google, Meta, and Amazon. Model answers are traced to a source or measured on real models. A free account saves progress and brings missed checks back after one day, then three, then seven.