Loading...
AI Engineering course · one course, beginner to advanced
Where the course starts: how production ML works (packaging, serving, pipelines, feature stores, monitoring), then how language models read text, with tokens, tokenizers, embeddings and search by meaning. Start here.
All 206 lessons plus every future lesson, with runnable code and enterprise diagrams. One payment, lifetime access.
The foundations of running machine learning in production: the production gap, packaging, serving, pipelines, feature stores, registries, monitoring, and RAG.
The ways a model that looked fine stops working once it is shipped, each one measured on two years of real bike-rental data: a test split that flatters the model, a world that changes after training, inputs computed differently at serving time, silent failures, drift you can watch before the answers arrive, and data that answers back.
Every step of the lifecycle, measured on three years of real electricity-market data: the baseline a model must beat, why two runs of the same code differ, the winner's curse when picking settings, the promotion gate, canary and shadow rollouts, when to retrain, stale pipeline pieces, rebuilding an old model exactly, and answers that arrive late.
How a language model reads text: tokens, tokenizers, token counts and cost, then embeddings, the numbers that stand for meaning, and how similarity search works on them. Every number measured with real tokenizers and real embedding models.
What a language model actually does, measured on a laptop: the odds it gives every next token, how sampling picks one, why reading a prompt is fast and writing is slow, attention, the cost of long prompts, caching, memory, context windows and chat templates.
Writing prompts you can measure, on a model running on your own laptop: the parts of a prompt, clear instructions, system and user messages, examples, output formats, quoting untrusted input, step-by-step reasoning, how small rewordings move the scores, and testing a prompt before you ship it.
When training a model beats writing a better prompt, measured on your own laptop: the cheapest baseline first, what training changes, LoRA, building and checking a training set, how many examples you need, overfitting, and testing a fine-tuned small model against prompted bigger ones on the same task.
206 lessons in total, across 3 levels
Data engineering for ML, retrieval and RAG in production, and running agents in production: pipelines, ingestion, versioning and data contracts, then chunking, retrieval and grounding, then tool calls, planning, memory and failure handling.
70 lessons6 chapters
Evaluating LLM systems, operating them in production, and securing them: error analysis, judges and eval statistics, fine-tuning vs RAG, inference and guardrails, then prompt injection and agent safety.
62 lessons3 chapters