Skip to main content
Course · 43 lessons

AI Engineer Roadmap.

A course from Python fundamentals to production AI systems: ML foundations, LLMs, embeddings, RAG, agents, MCP and real-world deployment.

This curriculum takes you from Python essentials through the full AI engineering stack — data handling, training intuition, working with foundation models, and shipping production-grade AI features that hold up under real traffic.

What you will learn

  • Python foundations required for data manipulation and model scripting.
  • Core ML concepts: supervised learning, validation that does not leak, and feature engineering.
  • How an LLM actually works underneath the API call — tokenization, transformer blocks, and the decoding parameters (temperature, top-k/top-p) that shape output.
  • Working with LLM APIs — prompt engineering, structured outputs, context design, tool use, and abstracting over multiple providers.
  • Building retrieval-augmented generation pipelines: embeddings, vector stores, hybrid search and reranking, and diagnosing why a RAG system gives the wrong answer.
  • Designing agents that plan, remember, and coordinate.
  • Connecting agents to tools and data over the Model Context Protocol (MCP), with a human approval step before destructive actions.
  • Evaluating, securing (beyond basic guardrails — prompt injection, tool abuse), deploying, and monitoring AI services in production.

This path is a curated on-ramp, not full /llm/ depth — model adaptation (LoRA/QLoRA, quantization), inference serving internals, multimodal inputs, and the pretraining→ alignment pipeline are real topics on the site but left for the LLM Engineering roadmap once you’re past the fundamentals here. MCP authorization, HTTP deployment and tracing are covered in the MCP Learning Roadmap.

The roadmap, step by step

1
python Step 1 of 43

Python Variables and Type System

See how Python variables work by running them: names as references, mutable vs immutable objects, dynamic typing, PEP 8 names and LEGB scope.

2
python Step 2 of 43

Python Data Types

Master Python's built-in data types — int, float, str, bool, None — and how to inspect and convert between them with real type conversion.

3
python Step 3 of 43

Python Functions and Parameters

Define Python functions with positional, keyword, default, *args and **kwargs parameters, see the mutable-default trap, lambdas and return values run.

4
python Step 4 of 43

Python Data Structures

Use Python lists, tuples, dicts, sets and comprehensions through small runnable programs, plus a real timing of list vs set membership tests.

5
python Step 5 of 43

Error Handling in Python

Handle exceptions with try/except/else/finally, chain and re-raise them, define custom exceptions, use except* and logging, all with real tracebacks.

6
python Step 6 of 43

Python Type Hints and Static Typing

Annotate Python code and run mypy on it: unions, narrowing, generics, Protocol, TypedDict and --strict, with the real errors each one catches.

7
numpy Step 7 of 43

Creating NumPy Arrays

Create NumPy arrays from lists, ranges, linspace, zeros, ones, full, eye and random generators, and see the float-step arange trap with real output.

8
pandas Step 8 of 43

Pandas DataFrames and Series

Select columns, pick rows with loc and iloc, filter with boolean masks and query(), add computed columns, and sort — every Pandas example run for real.

9
ml Step 9 of 43

Introduction to Machine Learning

Understand the core concepts of machine learning — supervised vs unsupervised learning, the bias-variance tradeoff, and the ML workflow.

10
ml Step 10 of 43

Supervised Learning Algorithms

Understand and apply the core supervised learning algorithms — when to use each, how they work, and how to tune them.

11
ml Step 11 of 43

Machine Learning Cross-Validation

Understand and implement k-fold, stratified, time-series, and nested cross-validation for reliable model evaluation.

12
ml Step 12 of 43

Feature Engineering Pipelines

Build robust, production-ready feature engineering pipelines that prevent data leakage, handle edge cases, and plug into scikit-learn.

13
ml Step 13 of 43

Model Interpretability

Explain model predictions with SHAP values, LIME, partial dependence plots, and permutation importance.

14
llm Step 14 of 43

Introduction to LLM Engineering

Understand how large language models work, learn the OpenAI API, and write your first production LLM application.

15
llm Step 15 of 43

How an LLM Actually Works

Trace a real prompt through tokenization, embeddings, transformer blocks, and next-token sampling using an actual small language model's real weights.

16
llm Step 16 of 43

Tokenization

How LLMs split text into tokens with byte-pair encoding, and why identical sentences cost very different token counts in different languages.

17
llm Step 17 of 43

Inference and Decoding

How a probability distribution becomes one chosen word — greedy decoding, temperature, top-k, top-p, and why temperature 0 isn't 'more certain'.

18
llm Step 18 of 43

Prompt Engineering

Write effective prompts using few-shot examples, chain-of-thought reasoning, role prompting, and structured output techniques.

19
llm Step 19 of 43

Structured Outputs in Production

Why 'return JSON only' isn't enough, and how schema validation plus a retry loop turns a flaky text model into a reliable data source.

20
llm Step 20 of 43

LLM Context Engineering

Fit system prompt, history, documents and tool results into a token budget, cite sources, chunk by heading and trim tool output, all as runnable Python.

21
llm Step 21 of 43

LLM Tool Use and Function Calling

Tool schemas, argument validation, the tool_use/tool_result loop, error recovery, parallel calls and approval gates, all run against a scripted model.

22
llm Step 22 of 43

LLM APIs and Provider Abstraction

A common interface over multiple LLM backends, with a real timeout-and-fallback router — so one provider's slow response doesn't become your outage.

23
llm Step 23 of 43

LLM Embeddings and Vector Search

What embeddings are, why keyword vectors miss meaning, cosine vs dot vs Euclidean, storage and search costs, choosing a model, and common uses.

24
llm Step 24 of 43

RAG — Retrieval-Augmented Generation

Build a production RAG pipeline: chunk documents, embed them, store in a vector database, retrieve on query, and generate grounded answers.

25
llm Step 25 of 43

RAG: Advanced Retrieval and Reranking

Hybrid search can rank the wrong document to the top — a real measured example, then the fixes: cross-encoder reranking and multi-query retrieval.

26
llm Step 26 of 43

RAG Failure Diagnosis

Why is my RAG system giving the wrong answer? Three real failure modes — bad chunking, stale documents, context overload — broken, then diagnosed.

27
ai-agents Step 27 of 43

Introduction to AI Agents

What makes an AI agent: tools, the ReAct loop of thought, action and observation, step caps that stop runaway loops, and memory that survives sessions.

28
ai-agents Step 28 of 43

AI Agent Planning Patterns

Plans as validated data, plan-then-execute with replanning on failure, parallel steps from a dependency graph, and checking results before acting.

29
ai-agents Step 29 of 43

AI Agent Memory Systems

Sliding windows that forget, fact summaries that don't, a SQLite store that persists across sessions, pinned memories, and deletion, all as runnable code.

30
ai-agents Step 30 of 43

RAG Agents

An agent loop that searches, reads, rephrases and searches again, keeps weak matches out, verifies its citations and says so when the docs don't answer.

31
ai-agents Step 31 of 43

Multi-Agent Systems

Orchestrators and subagents in runnable Python: parallel fan-out, partial failures, parsing a model's plan safely, and approval gates for risky tool calls.

32
mcp Step 32 of 43

MCP: Your First Server and Client

Build a Model Context Protocol server with one tool in Python, call it from a scripted client over stdio and in memory, and see why stdout is off limits.

33
mcp Step 33 of 43

MCP Tools: Schemas, Results and Errors

See the inputSchema MCP builds from Python type hints, return structured output, and tell tool errors from protocol errors, all with real printed results.

34
mcp Step 34 of 43

MCP Resources and URI Templates

Expose data as MCP resources: static URIs, {name} templates, MIME types and base64 blobs, plus what a client gets for a missing URI, all with real output.

35
ai-agents Step 35 of 43

An Agent Loop Over MCP Tools, With Human Approval

Build an agent loop on MCP: discover tools, validate a plan, run read-only tools freely, gate destructive ones on approval, and use several MCP servers.

36
llm Step 36 of 43

LLM Evaluation

Why overlap metrics mislead, assertion-based graders, judge calibration with kappa, retrieval recall and MRR, and a regression gate, all run in Python.

37
llm Step 37 of 43

LLM Safety and Guardrails

Input screening, escaped delimiters, canary tokens, PII redaction, and per-user rate limits for LLM apps, including what each guardrail misses, all run.

38
llm Step 38 of 43

LLM Security Beyond Guardrails

Indirect prompt injection, excessive agency, and insecure output handling — the OWASP-style risks that survive input/output guardrails.

39
mcp Step 39 of 43

MCP Server Security

Defend MCP servers and hosts: contain file paths, keep resource checks on, audit and pin tool descriptions, distrust annotations, and guard credentials.

40
llm Step 40 of 43

LLM Production Patterns

Retries with backoff and jitter, a SQLite response cache, token cost math, model routing, and a circuit breaker around LLM API calls, all run for real.

41
ai-agents Step 41 of 43

Agent Evaluation

Grade agents on outcome and process, measure pass rates over repeated runs, test safety under pressure, and audit logged trajectories.

42
mlops Step 42 of 43

Model Serving and Deployment

Serve a model with FastAPI and pydantic, load it once from an MLflow alias, batch-score Parquet in chunks, and gate promotion and rollback on a holdout.

43
mlops Step 43 of 43

MLOps Model Monitoring

Detect data drift with KS tests and PSI, avoid the large-sample p-value trap, see concept drift that inputs can't show, and turn checks into alerts.