A practical 40-lesson follow-on to "Machine Learning from Zero to Advanced". Go from how GPT models are trained and sampled, through prompting, embeddings, vector search, RAG and fine-tuning, to tool-using agents, MCP and production serving. It ends with four capstone projects and a roadmap. Every lesson gives intuition, math, a worked numeric example and runnable provider-neutral Python.
Aulas do Curso
(40 vídeos)
0 de 40 aulas concluídas
0%
Progresso salvo somente neste navegador.
0
How GPT Is Trained: Next-Token Prediction Explained with Math and Code
Every GPT model starts with one simple game: predict the next token. Here is how that game becomes a chatbot.
?? Read t…
1
Scaling Laws Explained: Why Bigger Models Keep Getting Better
Why does making a language model bigger keep making it better, and so predictably? In this lesson we unpack scaling laws…
2
Sampling Explained: Greedy, Temperature, Top-k, Top-p (With Numbers)
Why does the same LLM sound robotic in one app and wildly creative in another? The answer is not in the model. It is in …
3
Context Windows and RoPE Explained: How LLMs Know Word Order
How does a Transformer know which word came first? We derive Rotary Position Embeddings (RoPE) from scratch, work a 2D e…
4
The KV Cache Explained: Why LLM Decoding Is Memory-Bound
Why does a GPU that can do hundreds of trillions of operations per second still crawl through text one token at a time? …
5
Why LLMs Hallucinate (and What Actually Helps)
Why does a language model confidently invent facts, names and citations? Because it is trained to produce plausible text…
6
Prompt Patterns and System Prompts: Programming in Plain English
Prompting is programming with natural language specs. Learn the patterns that make LLM output reliable.
?? Read this le…
7
Few-Shot Prompting and In-Context Learning Explained (With Code)
What if you could teach a language model a brand new task without changing a single weight? That is in-context learning,…
8
Chain-of-Thought and Reasoning Models: How LLMs Learn to Think
Why does asking a model to think step by step change its answers? This lesson explains chain-of-thought prompting, self-…
9
Structured Output and Evaluating Prompts Like Real Software
Your LLM works in the demo, then breaks your parser in production. Here is how to fix both problems.
?? Read this lesso…
10
Sentence Embeddings and Contrastive Training: How Meaning Becomes Math
How does a computer know that "How do I reset my password?" and "I forgot my login" mean the same thing? Sentence embedd…
11
Cosine vs Dot Product: How Exact Vector Search Really Works
Your vector search can return the wrong results even when the embedding model is perfect. The cause is often a mismatch …
12
ANN Indexes Explained: HNSW, IVF and Product Quantization
How do vector databases search millions of embeddings in milliseconds? Learn how HNSW graphs, IVF clustering and product…
13
Vector Databases in Practice: Filtering, Sizing, and Real Code
An ANN index is not a database. Learn what a vector database really adds, and how to use one well.
?? Read this lesson …
14
RAG Explained: How to Ground LLMs in Your Own Data
What if your LLM could look up the answer instead of guessing it?
?? Read this lesson as an article, chapter by chapter…
15
Chunking Strategies for RAG: How to Split Documents Right
Your RAG system can only find what you chunked well. Learn fixed-size, recursive, structure-aware, semantic and parent-d…
16
Retrieval and Query Rewriting: Fix RAG Before It Fails
Your RAG bot has the right document and still answers wrong. The problem is often the query, not the index.
?? Read thi…
17
Reranking with Cross-Encoders: The Cheap Trick That Fixes RAG Search
Your vector search finds roughly the right documents, but the best one is stuck at rank 7. Cross-encoder reranking fixes…
18
Hybrid Search Explained: BM25 + Dense Vectors with Reciprocal Rank Fusion
Your vector search just missed an exact error code. Here is why, and how hybrid search fixes it.
?? Read this lesson as…
19
Evaluating RAG: Metrics, Test Sets and 8 Failure Modes
Your RAG demo looks great, so why does it fail in production? Because you are probably judging it by eyeballing a few an…
20
Supervised Fine-Tuning Explained: Chat Templates, Loss Masking and Code
How does a model that only predicts the next token become an assistant that follows instructions? In this lesson we buil…
21
LoRA Math Explained: Low-Rank Adaptation Derived From Scratch
Fine-tune a huge model by training less than half a percent of one weight matrix? Here is exactly why LoRA works.
?? Re…
22
QLoRA Explained: Fine-Tune a 7B Model on One GPU with 4-Bit Quantization
A 7 billion parameter model needs over 100 gigabytes to fine-tune the normal way. QLoRA does it in a fraction of that. H…
23
RLHF Explained: Reward Models, KL Penalty and PPO Step by Step
How does a raw language model become a helpful assistant? Reinforcement learning from human feedback, explained with rea…
24
DPO Derived: Preference Optimization Without RL
What if you could get the benefits of RLHF with a single, simple classification-style loss? In this lesson we derive Dir…
25
Fine-Tune vs RAG vs Prompt: How to Decide (A Practical Framework)
Should you prompt, retrieve, or fine-tune? Most teams pick the wrong tool first. This lesson gives you a clear decision …
26
Tool and Function Calling: How LLMs Actually Use Tools
A language model can only write text, so how does it check the weather, do exact math, or call your API? The answer is t…
27
The ReAct Loop, Planning and Reflection: How AI Agents Actually Think
Build an AI agent from scratch: the ReAct loop explained, plus planning, reflection and the failure modes that sink real…
28
Agent Memory Explained: Short-Term, Long-Term and Context Management
Your agent forgets everything the moment the context window closes. Here is how to fix that.
?? Read this lesson as an …
29
Multi-Agent Systems: When They Help and When They Hurt
More agents is not automatically better. Learn when splitting work across agents pays off, and when one agent wins.
?? …
30
Model Context Protocol (MCP) Explained: One Standard for AI Tools
Every AI app needs tools, and every tool needs every app. MCP is the open protocol that turns that mess into one clean s…
31
Latency and Cost Math for LLM Applications: Predict Your Bill
Your LLM demo costs a fraction of a cent. Your production bill says otherwise. Learn the math that explains why.
?? Rea…
32
Quantization Explained: INT8, INT4, GPTQ and AWQ for LLMs
How do you fit a 70 billion parameter model into a fraction of the memory without wrecking it? Quantization.
?? Read th…
33
Serving LLMs with vLLM: Continuous Batching and PagedAttention Explained
Why does the same GPU serve several times more users with vLLM? Two ideas: continuous batching and PagedAttention.
?? R…
34
LLM Guardrails and Prompt Injection: Securing and Monitoring AI in Production
Your LLM app passed every eval, so why can one pasted web page make it leak data?
?? Read this lesson as an article, ch…
35
Build a Document Q&A RAG Bot: Hybrid Search, Reranking, Citations, Evals
Most RAG demos work on three hand-picked questions and fall apart on the fourth. In this capstone we build a document Q&…
36
Capstone 2: Build a Repo-Aware Coding Assistant (Python, Sandboxed, Tested)
What does it take to build a coding assistant that edits your code and proves it works? In this capstone we build one fr…
37
Capstone: Fine-Tune a Small Open Model End to End (QLoRA + TRL)
Fine-tune a small open model for a narrow task, the right way: baseline first, clean data, QLoRA with TRL, honest evalua…
38
Capstone 4: Deploy a Tool-Using AI Agent to Production (FastAPI, Docker, Guardrails)
Build and ship a production-style tool-using agent: tool calling, memory, step and cost budgets, guardrails, human appro…
39
LLM Engineering Roadmap: What You Learned and Where to Go Next
You finished the course. Now what? This final lesson turns forty lessons into one mental map, from foundations to produc…