5 projects to become an AI engineer (no CS degree) ⚡
You don't become an AI engineer by watching tutorials, you become one by shipping things that work. These are the 5 projects that map to what actually show
⚡ TL;DR
- You don't become an AI engineer by watching tutorials, you become one by shipping things that work.
- These are the 5 projects that map to what actually shows up in real AI engineering job posts, in the order I'd build them.
- Every tutorial is free and first party, and you can build all of them with free accounts (the fine tuning one just needs a free Google Colab GPU).
BUILD THESE
The 5 projects (beginner to advanced)
1. Build a RAG chatbot over your own docs 📄
- 🧩 Stack: LangChain + Chroma or FAISS (vector database) + Claude API.
- 🔗 Free tutorial: https://huggingface.co/learn/cookbook/en/rag_zephyr_langchain
- 💡 Why it matters: retrieval augmented generation (chunk, embed, store, retrieve, generate) is THE foundational pattern, and the single most in demand skill in AI engineering job posts. You build a bot that answers questions from your own PDFs.
2. Build your own MCP server 🔌
- 🧩 Stack: Model Context Protocol + Python + the official MCP SDK + Claude Desktop.
- 🔗 Free tutorial (official): https://modelcontextprotocol.io/docs/develop/build-server
- 💡 Why it matters: MCP is how modern agents get tools. Building a server that gives Claude a new capability (read your files, call an API) proves you understand the current industry standard, not just prompting.
3. Build a tool-using AI agent 🤖
- 🧩 Stack: Claude API (tool use and function calling) + Python + the Anthropic SDK.
- 🔗 Free tutorial (official): https://platform.claude.com/docs/en/agents-and-tools/tool-use/build-a-tool-using-agent
- 💡 Why it matters: you hand write the agentic loop (the model decides when to call your functions and loops until the task is done), so you actually understand what an agent is under the hood instead of copying a template.
4. Fine-tune a small model on your own data 🎛️
- 🧩 Stack: Hugging Face Transformers + trl or Unsloth + a free Google Colab or Kaggle GPU.
- 🔗 Free tutorial: https://huggingface.co/blog/dvgodoy/fine-tuning-llm-hugging-face (Unsloth's free Colab notebooks are a lighter alternative: https://unsloth.ai/docs/get-started/fine-tuning-llms-guide)
- 💡 Why it matters: fine tuning a tiny open model (like Phi-3 Mini) for a narrow task takes you past prompting into shaping model weights, the line that separates prompt user from AI engineer.
5. Ship a multi-agent workflow ⚙️
- 🧩 Stack: n8n (AI Agent node) for the low code path, or LangGraph for the code path, plus Claude or OpenAI and real tools and APIs.
- 🔗 Free tutorial (official): https://docs.n8n.io/advanced-ai/intro-tutorial/
- 💡 Why it matters: moving from a notebook demo to a deployed, always on system (an orchestrator delegating to sub agents, or a workflow triggered by email or webhook) is the engineer, not tinkerer line.
USE IT WELL
How to make these count 📌
- 📌 Put every project on GitHub with a short README (what it does, the stack, a screenshot or a 20 second demo gif). The README is what recruiters actually read.
- 📌 Post a build in public clip of each one, so you get portfolio and content at the same time.
- 📌 Do them in order. Each one reuses skills from the last, so by project 5 the advanced one won't feel advanced.