NLP to LLMs Build the Understanding That Lasts

What you'll learn

---

Frame any text problem as a standard task type, and ship a TF-IDF baseline that earns its keep
Trace one attention head by hand, with real shapes, until the mechanism stops being a metaphor
Diagram a full transformer block from memory: multi-head, feed-forward, residuals, positions
Narrate the GPT lineage as a series of specific bets, and explain why chat models behave unlike base models
Control generation on purpose with temperature and top-p, and get reliable JSON and tool calls
Estimate and cut the cost of an LLM feature before you build it, using context and latency budgets
Build a RAG system and debug retrieval with recall checks rather than vibes
Run a real LoRA fine-tune on a laptop CPU, verify the gain, and defend the build decision with your own numbers
Requirements

---

Comfortable Python (you can read a class and a loop)
Some exposure to machine learning basics helps but is not required
No GPU, no paid API key, no cloud account. Every lab runs on a laptop CPU
Description

---

This course contains the use of artificial intelligence.
You can call an LLM API today. The question this course answers is what to do when the API is not the right answer.
Most LLM courses teach you to use a product. This one rebuilds the understanding underneath it, in the order the ideas actually developed: text as sparse counts, then dense meaning, then attention, then the models that scale it, then everything we have learned about making them behave.
The labs are the point, and they all run on your laptop. You will trace self-attention by hand on a toy example, with shapes, before you ever see the library call. The LoRA layer in module 7 is written directly against torch rather than pulled from a package, because the whole point is to see the low-rank update. The ROUGE implementation in module 8 is thirty lines of Counter. Nothing here is a black box you are asked to trust.
Every number in this course was measured, not estimated. The module 7 fine-tune takes 233 seconds and the capstone's four-system comparison about six minutes, both on one laptop CPU with sixteen torch threads. Timings are machine specific and labelled that way in the lessons; accuracies reproduce at the stated seeds.
The capstone makes you choose and defend. Same task, classical approach against the LLM approach, measured on cost and quality, and you write the decision memo. That memo is the artifact you show an employer, and it is worth more than another notebook with a high accuracy number in it.
The course also covers what most curricula skip: where models fail systematically, how to stand up a minimal eval before shipping, and how to threat-model prompt injection and leakage at practitioner level.
Module 1 is free. Watch it and you will know whether this is the right depth for you.
The other three OfferLab courses: Cracking the US Data Science & ML Interview · The US Tech Interview: Behavioral, Communication, Offer · The ML System Design Interview. Each stands on its own; take them in whatever order fits what you are preparing for.
Who this course is for

---

Engineers and data scientists who use LLMs daily and want to know what is actually happening
ML practitioners who learned the classical stack and skipped the transformer era
Anyone who can call an API but cannot yet say when to fine-tune instead
Not for: people who want a no-code prompt-tips course, or researchers who need paper-level depth
Homepage

---