Home/Roadmaps/AI Engineer
Roadmap · Updated May 2026

The AI Engineer trek

From Python fundamentals to production LLM systems. Models, RAG, evals, agents, and shipping real applications.

Stages
14
Estimated time
6 months
Level
Beginner → Advanced
Maintained by
3 practitioners
01
Stage 01

Python foundations

Idiomatic Python — types, comprehensions, generators, dataclasses, async. Enough fluency that you stop fighting the language and start expressing ideas in it.

PythonToolingBeginner
02
Stage 02

Math & statistics

Just enough linear algebra, calculus, and probability to read papers and reason about model behavior. We skip what you don't need.

MathStatsBeginner
03
Stage 03

Data manipulation

NumPy, pandas, Polars. Slicing, joining, reshaping, plotting. Skills that pay rent every single day in this field.

NumPyPandasPolars
04
Stage 04

Classical machine learning

Linear/logistic regression, trees, ensembles, clustering. Build the intuitions that make every later stage make sense — and learn what to reach for when an LLM is the wrong hammer.

scikit-learnMLIntermediate
05
Stage 05

Deep learning

PyTorch from autograd up. Build CNNs and RNNs from scratch before you ever import a pretrained model. The mechanics are the moat.

PyTorchDeep LearningIntermediate
06
Stage 06

NLP & transformers

Tokenization, embeddings, attention, the transformer block. Implement a tiny transformer from scratch — the rest of the field will feel inevitable afterwards.

NLPTransformersHuggingFace
07
Stage 07

LLM fundamentals

How modern LLMs are pretrained, what RLHF actually does, why context length matters, and the operational shape of an inference call.

LLMsRLHFInference
08
Stage 08

Prompting & evaluation

Structured prompting, output formats, and — more importantly — building eval harnesses that catch regressions before users do.

PromptingEvalsIntermediate
09
Stage 09

RAG systems

Embeddings, vector stores, chunking, hybrid retrieval, reranking, citation extraction. Build one end-to-end and break it deliberately.

EmbeddingsVector DBReranking
10
Stage 10

Fine-tuning

When fine-tuning helps and when it doesn't. LoRA, QLoRA, full SFT, preference optimization. Run a real fine-tune on a real dataset.

LoRAQLoRAFine-tuning
11
Stage 11

Agents & tool-use

Function calling, planner/executor loops, MCP, multi-step reasoning, and the failure modes that ship most agent demos straight to the graveyard.

AgentsTool-useMCP
12
Stage 12

Deployment & serving

vLLM, batching, quantization, autoscaling, cost models. Make an LLM-powered service that doesn't fall over at 10x traffic.

vLLMServingInfra
13
Stage 13

Observability & safety

Traces, structured logs, online evals, prompt injection defense, output filtering, and the on-call playbook for an AI system.

ObservabilitySafetyOn-call
14
Stage 14

Capstone — ship it

Build, deploy, evaluate, and write up an end-to-end AI system of your choosing. Open-source it. Get one practitioner review. This is the artifact you put on the top of your résumé.

CapstonePortfolioAdvanced

Trek complete. What's next?

You've walked the full roadmap. Now ship the capstone, write about it, and share the path with the next engineer who needs it.

Read the blogExplore more roadmaps