Learn GenAI engineering from zero to senior.

291 steps of 10-15 minutes, each with practice and a check. You finish with 12 systems you built yourself.

Step 2, live: how a model reads your text

0 tokens0 words

A rough split for the demo. Each model has its own tokenizer, and step 2 counts them exactly.

One start, two roads

Steps 0-31 need no code. Then developers go on to engineering, and business readers take a 20-step no-code branch that rejoins where AI meets marketing, sales, content and video.

0-31 Foundations, no code 32-74 Engineering basics 75-137 Harnesses, agents, skills 138-170 RAG, memory, multimodal 171-212 Production 213-232 Models in depth 233-258 AI across business functions 259-290 AI in business and senior practice n01-n20 No-code branch rejoins at step 233

12 systems you will have built

Each block ends with a capstone. Together they are your portfolio.

  1. step 31A prompt pack for a real process at your work
  2. step 48A CLI chat that shows what every answer costs
  3. step 66An invoice extraction service with validated output
  4. step 100Your own agent harness: tools, sandbox, approvals, memory
  5. step 109An MCP server in front of an internal API
  6. step 137A team of research agents with budgets and tracing
  7. step 154A company docs assistant that respects access rights
  8. step 166A voice agent that books appointments
  9. step 176A proof that a new prompt is better, with a significance check
  10. step 232A self-hosted model behind an OpenAI-compatible API, load-tested
  11. step 258A week of one department automated: hiring, HR, finance or legal
  12. step 290A final project on a real business problem, from spec to launch

Syllabus

Every step, in order. Steps 0-20 are free.

Part 1 Foundations, no code 0-31

What GenAI is 0-9
  1. 0What an LLM is, in one page free
  2. 1Next-token prediction free
  3. 2Tokens and the tokenizer free
  4. 3The context window free
  5. 4Training, fine-tuning, inference free
  6. 5The model landscape: open and closed weights free
  7. 6Why models hallucinate free
  8. 7Reasoning models and fast models free
  9. 8What a task costs: input and output tokens free
  10. 9Choosing a model for a job free
Working with AI every day 10-17
  1. 10Chat apps and their settings free
  2. 11Projects, custom instructions, memory free
  3. 12Files, images and long documents free
  4. 13Web search, citations, checking the answer free
  5. 14What never goes into a chat free
  6. 15Spreadsheets and data analysis with AI free
  7. 16Writing with AI without sounding like AI free
  8. 17Case: one workday with an AI assistant free
Prompting 18-31
  1. 18Anatomy of a prompt free
  2. 19Task, context, constraints free
  3. 20Few-shot: teaching by example free
  4. 21Structure with XML tags and sections
  5. 22Output format and length
  6. 23System prompts and roles
  7. 24Letting the model think
  8. 25Prompt chains: split the task
  9. 26Prompting over long documents
  10. 27Iterating against test cases
  11. 28Meta-prompting: the model writes the prompt
  12. 29Templates, variables, a prompt library
  13. 30Debugging a prompt that fails
  14. 31Capstone: a prompt pack for a real process

No-code branch: AI automation for business n01-n20

Opens after step 31. Rejoins the main path at step 233.

20 steps n01-n20
  1. n01Thinking in automations: trigger, steps, result
  2. n02n8n: install and first workflow
  3. n03Triggers and webhooks
  4. n04The LLM node
  5. n05Pulling fields out of emails
  6. n06Classify and route tickets
  7. n07Documents in, summaries out
  8. n08Sheets and CRM integration
  9. n09Errors and retries
  10. n10Human approval before sending
  11. n11The AI agent node with tools
  12. n12RAG without code: a vector store node
  13. n13Internal assistants: custom GPTs and Claude Projects
  14. n14MCP connectors in desktop apps
  15. n15Meeting notes on autopilot
  16. n16Costs and limits in no-code
  17. n17Data safety in no-code tools
  18. n18n8n, Make, Zapier: which one
  19. n19When to hand off to an engineer
  20. n20Capstone: one process automated end to end

Part 2 Engineering basics 32-74

Python for AI work 32-38
  1. 32Setup: uv, virtual envs, .env and secrets
  2. 33HTTP and JSON from Python
  3. 34async and concurrent calls
  4. 35Typed data with Pydantic
  5. 36A small CLI tool
  6. 37Tests with pytest
  7. 38Notebooks for exploring, scripts for shipping
First API calls 39-48
  1. 39API keys and the SDKs
  2. 40The messages API
  3. 41System prompt through the API
  4. 42temperature, top_p, max_tokens
  5. 43Streaming responses
  6. 44Multi-turn state
  7. 45Errors, rate limits, retries with backoff
  8. 46One interface, three vendors
  9. 47Local models with Ollama
  10. 48Capstone: CLI chat with a cost meter
Token economy 49-56
  1. 49Count tokens before you send
  2. 50Prompt caching
  3. 51Batch API for offline work
  4. 52Model routing: cheap first, expensive on demand
  5. 53Trimming and summarizing history
  6. 54Controlling output length
  7. 55Response and semantic caches
  8. 56Case: cut a monthly bill by 60%
Structured output and tool use 57-66
  1. 57Structured outputs and JSON schema
  2. 58Validate and repair with Pydantic
  3. 59Extracting fields from documents
  4. 60Classification and routing
  5. 61Function calling basics
  6. 62Designing a tool schema
  7. 63Parallel tool calls
  8. 64When a tool fails
  9. 65The agent loop, by hand
  10. 66Capstone: an invoice extraction service
Under the hood 67-74
  1. 67Embeddings and vector space
  2. 68Attention and the transformer, plainly
  3. 69Tokenizers across languages and what it costs
  4. 70Sampling: logits, temperature, top-k
  5. 71KV cache and why long context is slow
  6. 72Thinking budgets and reasoning tokens
  7. 73Context rot and needle tests
  8. 74How images and audio become tokens

Part 3 Harnesses, agents, skills 75-137

What a harness is 75-82
  1. 75A harness: loop, tools, context, permissions
  2. 76Inside Claude Code, Codex, Cursor, Gemini CLI
  3. 77Project memory: CLAUDE.md and AGENTS.md
  4. 78Permissions and sandboxing
  5. 79Hooks
  6. 80Slash commands and custom commands
  7. 81Context management: compaction and subagents
  8. 82Plan, build, review with a coding agent
Ready-made harnesses at work 83-90
  1. 83A real feature with Claude Code
  2. 84The same feature with Codex and Cursor
  3. 85Headless mode and scripts
  4. 86Agent code review on pull requests
  5. 87An agent inside GitHub Actions
  6. 88Worktrees and parallel sessions
  7. 89How agents fail: loops, overreach, fake success
  8. 90Measuring what the agent saved
Write your own harness 91-100
  1. 91A minimal agent loop in 100 lines
  2. 92A tool registry
  3. 93File and shell tools in a sandbox
  4. 94Permissions and approvals
  5. 95A context window manager
  6. 96Memory: scratchpad and long-term notes
  7. 97A planning and todo tool
  8. 98A terminal UI with streaming
  9. 99Agent SDKs: Claude Agent SDK, OpenAI Agents SDK
  10. 100Capstone: your own ops agent
MCP 101-109
  1. 101What MCP is and why it exists
  2. 102Using existing MCP servers
  3. 103Your first MCP server in Python
  4. 104Resources and prompts
  5. 105Remote MCP over HTTP with OAuth
  6. 106Testing with the MCP inspector
  7. 107MCP security: tool poisoning and scopes
  8. 108A2A: agents talking to agents
  9. 109Capstone: an MCP server for an internal API
Skills 110-115
  1. 110What a skill is: loading only what's needed
  2. 111Anatomy of SKILL.md
  3. 112Write a skill for your team
  4. 113Skills with scripts and reference files
  5. 114Does it trigger? Testing descriptions
  6. 115Plugins: packaging and sharing
Automation tools in demand 116-125
  1. 116A PDF and scan processing pipeline
  2. 117Email triage agent
  3. 118Support bot with a human handoff
  4. 119Meeting transcript to CRM
  5. 120Scraping plus LLM extraction
  6. 121Generated reports from raw data
  7. 122Browser automation with Playwright and an LLM
  8. 123Computer use
  9. 124Scheduled agents
  10. 125Slack and Telegram bots
Multi-agent workflows 126-137
  1. 126When more agents pay off, and when they don't
  2. 127Patterns: chain, router, parallel, orchestrator, evaluator
  3. 128Orchestrator and workers, by hand
  4. 129Handoffs between agents
  5. 130Shared state
  6. 131LangGraph
  7. 132Frameworks side by side: LangGraph, CrewAI, Agents SDK, subagents
  8. 133Durable execution with queues and Temporal
  9. 134Human-in-the-loop checkpoints
  10. 135Budgets, timeouts, failure handling
  11. 136Tracing a multi-agent run
  12. 137Capstone: a research agent team

Part 4 RAG, memory, multimodal 138-170

RAG 138-154
  1. 138RAG, long context or fine-tuning: how to choose
  2. 139Choosing an embedding model
  3. 140Chunking strategies
  4. 141Vector stores: pgvector and Qdrant
  5. 142The ingestion pipeline
  6. 143Top-k and metadata filters
  7. 144Hybrid search: BM25 plus vectors
  8. 145Reranking
  9. 146Query rewriting and HyDE
  10. 147Citations and grounding
  11. 148Contextual retrieval
  12. 149Measuring retrieval: recall@k, MRR
  13. 150Measuring answers: faithfulness and relevance
  14. 151Knowledge graphs and GraphRAG
  15. 152Agentic RAG
  16. 153RAG at scale: updates, deletes, access rights, tenants
  17. 154Capstone: a company docs assistant with access control
Memory and context engineering 155-158
  1. 155Types of agent memory
  2. 156Memory stores and what to forget
  3. 157Context engineering: what goes into the window
  4. 158Case: an assistant that remembers a customer
Multimodal 159-166
  1. 159Vision: screenshots, charts, photos
  2. 160Document AI: tables and forms in PDFs
  3. 161Speech to text
  4. 162Text to speech and realtime voice
  5. 163Image generation inside a product
  6. 164Video understanding
  7. 165Multimodal RAG
  8. 166Capstone: a voice agent that books appointments
AI product UX 167-170
  1. 167Streaming interfaces that feel fast
  2. 168Showing sources and confidence
  3. 169Collecting feedback inside the product
  4. 170Telling users the model can be wrong

Part 5 Production 171-212

Measuring quality, no math required 171-176
  1. 171Baselines: what you compare against
  2. 172Accuracy, precision, recall, F1 as ready formulas
  3. 173The confusion matrix, read in plain words
  4. 174Train, test and why a model can look better than it is
  5. 175Is the difference real? Significance with a ready calculator
  6. 176Case: prove the new prompt is better
Data for AI 177-180
  1. 177Synthetic data: generating test and training sets
  2. 178Data quality: duplicates, leaks, label noise
  3. 179A data pipeline for evals and training
  4. 180Versioning datasets
Evals 181-188
  1. 181Why evals come before features
  2. 182Building a golden dataset
  3. 183Code-based graders
  4. 184LLM-as-judge and its biases
  5. 185Human review and labeling
  6. 186Regression evals in CI
  7. 187Evaluating agent trajectories
  8. 188Online evals and A/B tests
LLMOps and observability 189-196
  1. 189Tracing with OpenTelemetry and Langfuse
  2. 190Prompt versioning and a registry
  3. 191Feature flags and canaries for prompts
  4. 192Latency: time to first token, parallel calls
  5. 193Fallbacks and provider outages
  6. 194Quotas and rate limits across tenants
  7. 195Deploying an LLM service: FastAPI, Docker, workers
  8. 196Cost monitoring and alerts
Security and law 197-204
  1. 197Prompt injection: direct and indirect
  2. 198Least privilege for tools
  3. 199Data leaks and PII redaction
  4. 200Jailbreaks and guardrails
  5. 201Red-teaming your own app
  6. 202Supply chain: models, packages, MCP servers
  7. 203GDPR for AI features
  8. 204EU AI Act: risk classes and duties
Incidents and documentation 205-208
  1. 205The model changed under you: catching regressions
  2. 206A hallucination incident: contain, roll back, explain
  3. 207Postmortems for AI systems
  4. 208Model cards and system documentation
Cloud and enterprise 209-212
  1. 209Bedrock, Vertex AI, Azure OpenAI
  2. 210Private networking and data residency
  3. 211SSO, roles and audit logs for AI features
  4. 212Procurement and security questionnaires

Part 6 Models in depth 213-232

Fine-tuning 213-224
  1. 213When fine-tuning beats prompting and RAG
  2. 214Preparing a training dataset
  3. 215Supervised fine-tuning through an API
  4. 216LoRA and QLoRA on an open model
  5. 217Preference tuning with DPO
  6. 218Distillation: big model teaches small
  7. 219Fine-tune against baseline: did it help?
  8. 220Fine-tuning embeddings for your domain
  9. 221Training a reranker on your own data
  10. 222RL with verifiable rewards: GRPO with a ready library
  11. 223Picking a model on your data, not on leaderboards
  12. 224Shipping an adapter
Inference and self-hosting 225-232
  1. 225Open models and their licenses
  2. 226Quantization: GGUF, AWQ, what you lose
  3. 227Serving with vLLM
  4. 228Batching, KV cache, throughput
  5. 229GPU sizing and cost: self-host or API
  6. 230Speculative decoding and small models
  7. 231Self-hosted embeddings and rerankers
  8. 232Capstone: an OpenAI-compatible endpoint under load

Part 7 AI across business functions 233-258

Marketing 233-237
  1. 233Market and competitor research with sources
  2. 234Customer insight from reviews, calls and tickets
  3. 235Ad copy and creative variants, tested
  4. 236An SEO and GEO content pipeline with a human editor
  5. 237Personalized email campaigns at scale
Sales 238-242
  1. 238Lead enrichment and scoring
  2. 239Personalized outreach that does not read as spam
  3. 240Call recordings into CRM notes and next steps
  4. 241Deal coaching from call transcripts
  5. 242Proposal and quote drafts from a template
Content creation 243-247
  1. 243Brand voice as a skill
  2. 244One long piece into ten formats
  3. 245Images for content: generation and editing
  4. 246Fact-checking and AI-tell review before publishing
  5. 247A content calendar agent
Video 248-253
  1. 248Editing by transcript: cut video like a document
  2. 249Subtitles, translation and dubbing
  3. 250Long video into short clips
  4. 251Generated video and avatars: rights, disclosure, limits
  5. 252Thumbnails, titles and descriptions
  6. 253A video pipeline from recording to publishing
Operations 254-258
  1. 254Recruiting: CV screening without bias traps
  2. 255HR: onboarding assistant and policy answers
  3. 256Finance: invoices, receipts, reconciliation
  4. 257Legal: contract review with a human in the loop
  5. 258Capstone: automate one department's week

Part 8 AI in business and senior practice 259-290

AI in business 259-268
  1. 259Finding processes worth automating
  2. 260Scoring feasibility
  3. 261ROI and total cost of ownership
  4. 262Build, buy or wait
  5. 263The discovery interview with a client
  6. 264Writing the proposal and scope
  7. 265Pilot design and success metrics
  8. 266Adoption: getting people to use it
  9. 267Pricing AI services
  10. 268Case: present a plan to stakeholders
AI system design 269-276
  1. 269A framework for designing AI systems
  2. 270Design: a support copilot
  3. 271Design: 1M documents a day
  4. 272Design: enterprise search
  5. 273Design: a coding agent platform
  6. 274SLOs for non-deterministic systems
  7. 275Ownership: who owns prompts, evals and models
  8. 276Mock: the AI system design interview
Working as a senior 277-280
  1. 277Estimating AI projects when the result is uncertain
  2. 278Explaining probabilistic results to leadership
  3. 279Mentoring and reviewing AI code
  4. 280Reading papers and filtering hype
AI Engineer interviews 281-284
  1. 281Live coding with an LLM API
  2. 282The take-home assignment
  3. 283Behavioral questions for AI roles
  4. 284Mock: the full interview loop
Final project 285-290
  1. 285Pick a real business problem
  2. 286Spec and evals first
  3. 287Build: agent, RAG and tools
  4. 288Harden: security, tracing, cost
  5. 289Launch and measure
  6. 290Portfolio write-up and demo

How practice works

Your key or a local model

Practice runs on your machine with an Anthropic, OpenAI or Gemini key, or a free local model through Ollama.

Checks that cost nothing

Every practice ships with recorded model responses, so the check runs offline. Live mode runs the same check on your key.

Cost in every step

From part 2 on, each step says what the technique costs in tokens and how to make it cheaper.

Scored by a metric

RAG, evals and fine-tuning steps score your work with recall, accuracy or faithfulness against a threshold.

Who it is for

Developers

You write code and want to work with LLMs at a senior level: agents, RAG, evals, fine-tuning, serving, system design.

Business side

You run operations, sales or a team. The no-code branch uses n8n and ends with one of your processes automated end to end.

FAQ

Do I need to know Python?

Not for steps 0-31 or the no-code branch. The engineering path has a short Python block (steps 32-38) for people who already code in another language.

Who pays for the tokens?

You use your own API key or a free local model. Each practice step shows its approximate cost before you run it.

Which models and tools?

Claude, GPT, Gemini and open models side by side. Claude Code, Codex and Cursor as ready-made harnesses, then you write your own.

What does senior mean here?

You build your own harness, eval pipeline and model serving, compare GPU cost with API cost, and design AI systems under load. Parts 5, 6 and 8 cover that, plus interviews and the senior role in a team.

Do I need math?

No. Where a formula matters (F1, significance, cosine similarity), the step gives it ready, explains it in words with a worked example, and code computes it. You never derive anything by hand.

Why 291 steps?

One step is one idea and one practice, 10-15 minutes. A short step fits into a weekday evening.

Will it go stale?

Model names, prices and APIs are checked against vendor docs on the day a step is written, and the step shows that date. Concepts and tools sit in separate parts of a step, so a new tool means a new example.

Is there AI for marketing, sales and video?

Yes, part 7 (steps 233-258): marketing, sales, content creation, video editing and operations such as hiring, HR, finance and legal. The no-code branch joins here too.

Languages?

English and Russian, mirrored step by step.

What does it cost?

Steps 0-20 are free. After that, a yearly plan or lifetime access. The waitlist hears about prices first.

Get the first steps when they open

One email when steps 0-20 go live.