AI Engineer Path
From data prep to LLMOps — complete production-grade engineering skills
Path goal
Finish able to own the full loop from data prep, model training/calling and evaluation to deployment and monitoring, reliably pushing ML or LLM apps to production.
Who it's for：Learners with Python and ML basics aiming to become AI/LLM engineers
-
Step 1: Engineering-grade Python
Move from writing scripts to writing engineering: virtual environments, tests and code structure, practiced through project tutorials.
- Official Python Tutorial
- Harvard CS50's Introduction to Programming with Python
- Project Based Learning Tutorial Collection
💡 Code that runs and code that's maintainable are different things — build good habits early.
-
Step 2: Math and machine learning
Build probability and statistics intuition and systematically study classic ML methods and metrics.
- Andrew Ng's Machine Learning Specialization
- Introduction to Statistical Learning Official Site
- scikit-learn User Guide
- StatQuest Statistics Fundamentals
- MIT 6.036: Introduction to Machine Learning
💡 Metrics matter more than models — understand what each one optimizes.
-
Step 3: Deep learning and NLP
Master deep learning basics and core NLP concepts, read the original Transformer paper, and understand pretrain-then-finetune.
- Stanford CS224n Natural Language Processing with Deep Learning
- Dive into Deep Learning (Chinese Edition)
- Hugging Face NLP Course
- The "Attention Is All You Need" Paper
- The BERT Paper
💡 Jay Alammar's illustrated Transformer is the best warm-up before the paper.
-
Step 4: LLM app development
Call mainstream LLM APIs fluently, understand in-context learning, prompting and multi-provider management, and unify routing with an API gateway.
- OpenAI API Official Docs
- Hugging Face LLM Course
- OpenRouter
- LiteLLM Documentation
- Prompt Engineering Guide (Chinese Edition)
💡 Production needs multi-model fallback and cost control — gateways like LiteLLM help.
-
Step 5: RAG systems engineering
Master retrieval, chunking, reranking and hybrid search, and quantify retrieval and generation quality with evaluation frameworks.
- LangChain RAG Tutorial
- Pinecone Learn
- Chroma Vector Database Documentation
- RAGAS Evaluation Framework Documentation
- RAG Survey Paper
💡 RAG quality lives in retrieval and context design, not in a bigger model.
-
Step 6: Agents and MCP
Understand agent tool calling and orchestration, master MCP, and wrap internal systems into standard tools for models.
- OpenAI Agents SDK Documentation
- MCP Servers Reference Repository
- CrewAI Documentation
- Agentic AI Learning Roadmap (Repository)
💡 MCP is the standardized path for wiring internal systems into AI — worth real investment.
-
Step 7: Fine-tuning and customization
Learn parameter-efficient methods like LoRA and fine-tune open models on your own data with LLaMA-Factory.
- LLaMA-Factory Fine-Tuning Framework
- LLM Fine-Tuning Datasets Collection
- Hugging Face Transformers
- Alibaba Qwen Open-Source Models
💡 Solve with prompting and RAG first; fine-tune only when you hit a real quality ceiling.
-
Step 8: Deployment and LLMOps
Serve models with vLLM or TGI, manage experiments and versions with MLflow and W&B, and set up monitoring and alerting.
- vLLM Official Documentation
- Hugging Face Text Generation Inference
- MLflow Official Documentation
- Made With ML Course
- Weights & Biases Documentation
- Full Stack Deep Learning
💡 Made With ML and Full Stack Deep Learning are the two best courses for this stage.
-
Step 9: Evaluation, security and a capstone
Build systematic evaluation and red-team awareness, and finish a capstone project following the AI engineer roadmap repo.
- OWASP Top 10 for LLM Applications
- AI Engineer Roadmap Repository
- Kaggle Competitions
- Build Your Own X
💡 End with one complete project — it's both proof of skill and your best calling card.