LLM Forge
Get access
16 modules · 33 hands-on notebooks

From tokens to agents.
Master modern LLM engineering.

A hands-on course from tokenization to production agents, retrieval, fine-tuning, alignment, evaluation, and observability. Read, get quizzed, and build it yourself in 33 runnable notebooks.

Read Module 1 freeSee pricing

Module 1 is free to read, no account needed.

View full curriculum ↓
Module 2 · Embeddings

2.3 Cosine similarity

Compare two vectors by the angle between them, not their length:

cos(θ) = (A·B) / (‖A‖‖B‖)

A score near 1 means the same direction — same meaning, regardless of magnitude.

Quick check

Why cosine similarity instead of Euclidean distance for embeddings?

Ignores vector magnitude, focuses on direction
It's faster to compute
Module 3 · Transformer Architecture

3.5 Scaled dot-product attention

With Q, K, and V in hand, the entire mechanism is one formula:

Attention(Q,K,V) = softmax(QKᵀ/√dₖ)·V

Divide by √dₖ to keep softmax in its sensitive, well-behaved zone...

Quick check

Why divide by √dₖ specifically?

Normalizes variance so softmax doesn't saturate
Makes computation faster
Module 6 · Retrieval-Augmented Generation

6.2 Retrieval before generation

Ground the model in your own documents instead of its training data:

query → embed → top-k chunks → prompt → LLM

The model only ever answers from what you hand it in context.

Quick check

What problem does RAG primarily solve?

Answering from private or up-to-date knowledge the model wasn't trained on
Making the model respond faster
Module 8 · Fine-Tuning & PEFT

8.4 LoRA: low-rank adaptation

Freeze the base weights, train a small low-rank update instead:

ΔW = B·A, rank r ≪ d

You update under 1% of parameters and still shift model behavior.

Quick check

Why does LoRA use two small matrices instead of one full-size one?

A low-rank product needs far fewer trainable parameters
It trains faster on a single GPU
Module 11 · Agentic AI

11.3 The tool-calling loop

The model doesn't act directly — it asks, you execute, and feed the result back:

think → call tool → observe → repeat

Each loop iteration gives the model fresh information to decide its next step.

Quick check

Who actually executes the tool call?

Your application code, not the model itself
The model runs it in a sandbox

What you'll build

Every concept lands in a notebook you run yourself, not slides you watch.

Build a search engine from scratch

Tokenization, TF-IDF, and embeddings, see exactly how raw text becomes numbers a model can reason with, before you ever call an API.

Ship a real RAG pipeline

Parse, chunk, embed, retrieve, and re-rank your own documents, then measure whether it's actually grounded with Faithfulness and Answer Relevancy.

Build an agent with real tools

Wire an LLM into the reason-act-observe loop, connect it to MCP, and evaluate whether it actually completes multi-step tasks.

Deploy it like production, not a demo

Fine-tune with PEFT, align with DPO, then trace, monitor, and optimize inference, quantization, batching, routing, so it survives real traffic.

What you can do afterwards

The gap between "I called an API once" and "I can design, evaluate, and ship a reliable LLM system" is where the leverage is. The tools move fast, but the fundamentals underneath them — tokenization, attention, retrieval, alignment, evaluation — are stable enough to learn properly, once, and reason from for years.

Explain what the model is actually doing

Tokenization, embeddings, attention — from first principles, with the maths worked through rather than waved at.

Build retrieval you can defend

Chunking, hybrid search, re-ranking, and the evaluation to show whether an answer is grounded or invented.

Ship an agent that completes real tasks

The reason-act-observe loop, tool calling, MCP, and the traces to tell you why a run went wrong.

Adapt and operate a model in production

PEFT fine-tuning, DPO alignment, quantization, batching, routing, monitoring, and drift detection.

Across 16 modules, 61 lessons, and 33 notebooks you run yourself.

The full curriculum

16 modules · 61 lessons · 33 notebooks, grouped into 6 themes.

Getting Started

Foundations

Working with LLMs

Retrieval

Adapting models

Production systems

How the course works

01

Read

Short, focused lessons with the formulas rendered like a proper textbook, real diagrams, and worked examples, not a slide deck.

02

Get quizzed

Every module ends in a multiple-choice check with instant feedback, dynamic, like the compliance training that actually made you remember something.

03

Build it

Download the notebook, run it locally or in Colab, and implement the thing you just read about with your own hands.

Sandro Zangiacomi teaching attention mechanisms at a blackboard

Who's teaching this

Sandro Zangiacomi

I teach AI the way I've built it. Before Stripe, I spent three years at Amazon Web Services as an Applied AI Engineer, architecting production GenAI systems, RAG pipelines, AI agents, and evaluation frameworks, across 10+ customer projects. I'm now a Solutions Architect at Stripe, advising AI-driven startups on how to turn LLM ideas into scalable, revenue-generating products.

I've also taught this material formally for years: I'm a guest lecturer at Efrei Paris, where I built and run a postgraduate course on Natural Language Processing & Chatbots, and a Visiting Teacher at Wuhan University of Technology in China, delivering that same course in person. I've also trained corporate leadership teams on generative AI fundamentals. This course distills everything from those classrooms into one place.

Connect on LinkedIn

What students say

From the postgraduate lectures this course is built from.

“This was an excellent class. The lectures were not only interesting but also broken down in a way that made everything easy to understand.”
Postgraduate student, Wuhan University of Technology
“The class atmosphere is pleasant and the teaching content is valuable. I'm totally satisfied with this course.”
Postgraduate student, Wuhan University of Technology

Written feedback collected at the end of the in-person course at Wuhan University of Technology. Quoted verbatim, published with names withheld.

Simple pricing. Everything included.

  • 16 modules, 61 lessons, full LaTeX-quality formulas and diagrams
  • 33 hands-on Jupyter notebooks (download or run in Google Colab)
  • Dynamic multiple-choice quizzes with instant feedback, every module
  • 5 click-through interactive walkthroughs of key mechanisms
  • Progress tracking across the whole course
  • Full access while subscribed, cancel anytime

€149/yr

Billed yearly, cancel anytimeSave €79 · 35%

Works out at €12.42/mo

Subscribe, €149/yr

Cancel anytime from your account — and Module 1 is free to read before you decide.

Secure checkout via Stripe. Questions? llmforge-support@agentmail.to

Want to read it first? Module 1 is free, no account needed.

Frequently asked questions

Do I need a deep learning background?+

No. You need to be comfortable with Python and the command line. Every concept, including the math, is built up from first principles, with worked examples before any formula.

Can I see the material before paying?+

Yes — Module 1 is free to read in full, with no account and no card. The lab notebooks for it need a free account so your progress can be saved.

How long does the course take?+

Most people move through it in 4–8 weeks at a few hours a week. Lessons are short (2–15 minutes each); the notebooks are where the real time goes.

What's the difference between monthly and annual?+

Exactly the same access. Monthly is €19/month; annual is €149/year, which works out at €12.42/mo — a saving of €79, or 35%. Either can be cancelled anytime.

Do I need my own compute or API keys?+

A laptop is enough for most notebooks. A few labs call an LLM API, the setup guide inside the course walks through getting free-tier keys, and every notebook also runs on Google Colab's free tier.

Is this course kept up to date?+

The fundamentals (tokenization, attention, retrieval, evaluation, alignment) are stable and don't chase every model release. Where the field is moving fast, we say so explicitly and point you to the primary sources.

Can I cancel?+

Cancel anytime from your account page and you won't be charged again. You keep access until the end of the period you've already paid for. Module 1 is free to read first, so you can judge the course before you subscribe. Any billing question, email llmforge-support@agentmail.to.

Do I get an invoice/receipt?+

Yes, Stripe emails a receipt automatically at checkout, and you can request a formatted invoice by replying to it.