Outcome
Explain model behavior, token costs, model trade-offs, structured outputs, embeddings, and tool-calling basics without hand-waving.
Foundations · Days 1-7
LLM Fundamentals for Engineers
Start with the mechanics behind modern language models from an engineering lens: tokens, model families, sampling, streaming, structured outputs, tool use, multimodal inputs, embeddings, and reasoning models.
Foundation 10 subtopics 7 daily blocks
Practice builds
Model comparison playgroundToken cost estimatorStructured output validator
Visual lessons
Watch the concepts move.
Short, diagram-led lessons that turn this roadmap into concrete engineering mental models.
2:36
How LLMs See Text: Tokens, Context Windows, and Cost
A diagram-led explanation of tokenization, shared context budgets, safe overflow policies, and how token usage becomes an operating cost.
Watch on YouTube ↗What to learn
Tokens, tokenization, context windows, and how they map to cost
Model families and trade-offs: Claude, GPT, Gemini, Llama, Mistral, Qwen, DeepSeek
Closed vs open-weight models, and when each makes sense
Sampling parameters: temperature, top_p, top_k, frequency and presence penalties
Streaming vs non-streaming responses
Structured outputs: JSON mode and schema-constrained generation
Function calling and tool use mechanics
Multimodal inputs: vision, audio, and PDFs
Embedding models and reranker models
Reasoning models vs standard models and when extended thinking helps
Daily study plan
Day 1: Learn tokens, context windows, pricing, and why long prompts become product costs.
Day 2: Compare closed and open-weight model families and write a simple model selection matrix.
Day 3: Experiment with temperature, top_p, and penalties on the same prompt.
Day 4: Build one streaming response endpoint and one non-streaming endpoint.
Day 5: Generate JSON with a schema and handle validation failure cleanly.
Day 6: Test function calling with one safe tool, like a weather or calculator stub.
Day 7: Summarize when to use embeddings, rerankers, multimodal inputs, and reasoning models.