🇷🇺 Русская версия | 🇺🇸 English version
🔥 Stop wasting time on useless AI courses.
This roadmap shows how to build real AI products in 2026 — step by step.
Don't know where to begin? Eliminate analysis paralysis:
- 🖥️ Interactive Dashboard — track your progress, steps, and projects
- 🟢 Beginner? → Start with the Beginner Path
- 🟡 Have coding experience? → Jump straight to the Developer Path
- 💰 Want to launch a product? → Go directly to the Money Path
Important: Do not skip the projects. They are your primary output. Build > Learn.
Goal: Understand the capabilities of LLMs and learn how to orchestrate them without writing complex code.
- LLM Basics: How tokens, context windows, and temperature work.
- Prompt Engineering & System Prompting: Understanding reasoning models (system prompts, managing Extended Thinking, contextual prompting).
- Tools: ChatGPT (o3-mini / GPT-4o / GPT-5), Claude (Claude 3.7 Sonnet / 3.5 Haiku, Artifacts / Claude Code CLI), Gemini (Gemini 2.5 Flash / Google AI Studio), v0.dev.
Why it matters: 👉 This is the foundation for automating marketing, SMM, and content agencies. Task description: Build a system that takes a single idea and transforms it into 5 different content formats (Summary, Telegram post, Twitter thread, email newsletter, video script). Success criteria: The system outputs ready-to-publish content that requires no manual editing.
📁 Starter template: projects/01-content-factory/
Goal: Create autonomous systems that can work with external data.
- Backend: Python (FastAPI) or Node.js.
- AI Frameworks & Protocols: PydanticAI, LangGraph, OpenAI SDK, Model Context Protocol (MCP).
- RAG & Vector Search: Vector Databases (Pinecone/Qdrant), pgvector (Supabase), hybrid search.
Why it matters: 👉 This is the baseline for corporate assistants and smart knowledge search tools within companies. Task description: A Telegram bot using RAG to answer questions strictly based on your files (PDFs/Links), coupled with a custom MCP server to connect the bot to your local file system or Notion. Success criteria: The bot does not hallucinate facts, referencing your documents and using MCP for access.
📁 Starter template: projects/02-telegram-mcp/
Goal: Turn code into a service that people are willing to pay for.
- UI & AI Integration: Next.js 15 + Tailwind v4 + v0.dev + Vercel AI SDK.
- Optimization & Routing: Outperforming routing (Semantic Routing) and prompt caching to optimize API costs.
- Payments & Auth: Stripe + Clerk.
Why it matters: 👉 This is your first working business in the cloud. Task description: A full-featured web application that solves a narrow B2B problem (e.g., an AI-powered legal contract generator). Success criteria: Working UI, user authentication, semantic request caching, and a basic Stripe payment integration.
📁 Starter template: projects/03-micro-saas/
Why it matters: 👉 Advanced multi-agent orchestration is the core pattern for complex 2026 AI systems. Task description: An autonomous loop with Planner, Executor, and Reviewer agents collaborating to solve tasks. Success criteria: Self-correcting feedback loop between Planner, Executor, and Reviewer until output passes validation.
📁 Starter template: projects/04-multi-agent-orchestrator/
Why it matters: 👉 Software engineering in 2026 has transitioned to autonomous CLI agents (Claude Code / OpenHands / Antigravity) capable of repository navigation, test execution, bug self-correction, and Git commits.
Task description: A complete engineering agent inside a sandboxed workspace: 6 tools (list_dir, read_file, write_file, run_command, git_action, run_tests), anti-path-traversal protection, self-correction on test failures, and session persistence to execution_trace.json.
Success criteria: The agent initializes Git, writes module code with unit tests, diagnoses failure tracebacks, applies self-correcting patches, verifies 100% test pass, and commits to Git.
📁 Starter template: projects/05-autonomous-agent-cli/
If you complete this curriculum:
- You will have 5 practical projects progressively engineered from working prototypes to production-ready grade.
- You will master the full lifecycle of modern AI engineering: from LLM generation and grounded RAG to FastMCP servers, vector-cached Micro-SaaS, and a flagship autonomous coding agent with Git and Pytest.
- You will learn how to architect autonomous systems with self-correction, thinking budget controls, and telemetry budgeting.
- You will be equipped to tackle real business challenges, take high-value client contracts, or build your own AI startup.
This is the journey of an AI Builder: from idea and sandbox to resilient production systems.
In 2026, those who build faster win over those who just know more. You do not need to:
- ❌ Study advanced mathematics for years.
- ❌ Take 50 theoretical courses in a row.
- ❌ Become a traditional ML researcher.
👉 Build first. Learn what's missing later.
Each of the 5 starter projects is designed with a dual-tier execution architecture:
- ⚡ Live LLM Mode (Production API): Connects to live frontier LLMs (OpenAI, Claude, Gemini), tool-calling APIs, and Stripe webhooks. Requires API keys in
.env. - 🧪 Guided Sandbox Demo (Zero-Dependency Walkthrough): Runs locally without any API keys (
--demo). Demonstrates agent loops, tool dispatch, test verification, and Git commits. Semantic search in demo mode uses TF-IDF cosine vector similarity to demonstrate mathematical principles cleanly without requiring gigabytes of dense neural model downloads.
- Frontier Reasoning: OpenAI GPT-5 / o3, Anthropic Claude 3.7 / 4 Sonnet, Google Gemini 3.8 / 3.7 Flash Thinking, DeepSeek-R1.
- Fast & High-Throughput: Gemini 2.5 Flash, GPT-4o-mini, Llama 3.3 70B (Groq LPU).
- Autonomous CLI Agents: Claude Code CLI, Cursor Agent, Antigravity, OpenHands, Qwen 2.5 Coder.
- Protocols: Model Context Protocol (MCP 2.0).
- Orchestration & AI SDK: LangGraph / PydanticAI / Vercel AI SDK 4+.
- Backend & Vector Stores: Python 3.12 / FastAPI, pgvector (Supabase), Qdrant.
- Frontend: React 19 / Next.js + Tailwind v4.
If this roadmap helped you save time:
- Give it a star ⭐
- Fork the repository
- Share it with a friend
- Contribute — see CONTRIBUTING.md
It helps grow the project and add new case studies.