AI Engineer Path

The line

Forty weeks, eight phases, about 520 hours. Every station is a week with its own lessons, videos, daily plan and self-check. Filled stations are weeks you've finished.

1
Prerequisites
Weeks 1–5
Five weeks to clear every prerequisite node. If you already program for a living, Python is a dialect, not a new skill, so move fast through syntax and slow down on Pydantic, asyncio and the maths. The two capstones (gradient descent in pure NumPy, and an EDA with five non-obvious findings) are the real exit tests.
Week 1Python core
Read and write Python without friction.
7 lessons
Week 2Python for services
The Python equivalents of a production backend toolkit.
8 lessonsBuilds: Port a small service to FastAPI + pytest + Docker
Week 3SQL · AI coding · DSA for AI
Clear three prerequisite nodes you mostly already hold.
6 lessons
Week 4Mathematics: linear algebra and calculus
Intuition, not rigour.
7 lessonsBuilds: Gradient descent in pure NumPy
Week 5Statistics and the data stack
Probabilistic literacy plus working NumPy and Pandas.
8 lessonsBuilds: EDA with five non-obvious findings
2
Machine Learning
Weeks 6–11
Don't skip this to reach LLMs faster. Every evaluation idea you will use on LLM systems (leakage, metric choice, held-out sets, baselines) is born here, and interviewers still test it. Gradient boosting remains the right answer for most tabular problems in industry.
Week 6ML fundamentals
Frame problems correctly and measure models honestly.
6 lessons
Week 7Feature engineering (1/2)
Turn raw, messy columns into model-ready features without leaking.
6 lessons
Week 8Feature engineering (2/2)
Create, select and compress features, and handle imbalance.
5 lessons
Week 9ML algorithms (1/2)
Know the classic supervised models well enough to choose and defend one.
6 lessons
Week 10ML algorithms (2/2)
Ensembles for prediction, and unsupervised methods for structure.
6 lessons
Week 11Advanced concepts + Project 1
Tune, ensemble and explain models, then ship one.
6 lessonsBuilds: Tabular ML pipeline, deployed
3
Other Modalities
Weeks 12–14
NLP is kept in full because it is the direct ancestor of the whole AI engineer track: TF-IDF and Word2Vec are the roots of embeddings and vector search. Computer vision is trimmed to what connects downstream (CLIP, ViTs, document understanding). Time series is half a day of awareness.
Week 12NLP (1/2): representing text
Turn text into numbers, from counting words to learned embeddings.
6 lessons
Week 13NLP (2/2): classic models + Project 2
Build classic NLP pipelines and a baseline you'll compare everything against.
6 lessonsBuilds: TF-IDF text classifier (the LLM baseline)
Week 14Computer vision (trimmed) + time series (awareness)
Just enough classical CV, then the multimodal models that matter now. Plus three time-series rules.
5 lessonsBuilds: Document understanding via a vision model
4
Deep Learning
Weeks 15–19
Five weeks that end at the hinge of the whole path: attention. RNNs are learned as a problem statement (why they lost), not as tools. If you are behind, slow down in week 19 rather than skim it.
Week 15Foundations of neural networks
Understand a neural network all the way down: forward pass, loss, backward pass, update.
5 lessons
Week 16Optimising a neural network
Make training fast, stable and well-generalising.
5 lessons
Week 17CNNs and transfer learning
Understand convolutional networks and, more importantly, transfer learning.
4 lessons
Week 18Sequence models: why RNNs lost
Learn RNNs as a problem statement, not as tools.
6 lessons
Week 19Seq2seq and attention
Understand the transformer block completely.
7 lessonsBuilds: Draw the transformer block from memoryHinge week: slow down hereNever cut
5
Practical Deep Learning
Weeks 20–22
The whole LLM ecosystem is PyTorch, so this is where theory turns into idiomatic training code. You will beat your own TF-IDF baseline with a fine-tuned BERT and measure by exactly how much. Week 22 is deliberately light: it is the buffer.
Week 20PyTorch
Write idiomatic PyTorch training code.
6 lessons
Week 21NLP with deep learning: Hugging Face
Use and fine-tune pretrained transformers with the Hugging Face stack.
5 lessonsBuilds: Fine-tuned BERT classifier vs your baseline
Week 22Unsupervised DL, deployment and review
Representation learning, a deployed demo, and a written review of everything so far.
6 lessonsLight week + buffer
6
Build a GPT
Weeks 23–25
An off-roadmap insertion, chosen deliberately. Understanding what sits beneath the abstraction is what lets you diagnose a RAG system returning garbage or an agent stuck in a loop. Type every line; watching is worthless. Short on time? The 'Let's build GPT' and tokenizer videos (about 4 hours) capture most of the value.
Week 23micrograd and makemore
Build backpropagation from nothing, then character-level language models.
5 lessonsOff-roadmap insertion
Week 24Let's build GPT
Implement attention from scratch and train a working GPT on your own corpus.
5 lessonsBuilds: Your own GPT, trained from scratchOff-roadmap insertion
Week 25Tokenizer, nanoGPT and multimodal models
Open the tokenizer black box, read production GPT code, and see how vision enters a language model.
6 lessonsBuilds: All from-scratch code on GitHubOff-roadmap insertion
7
AI Engineer Track
Weeks 26–37
30% of the plan and the reason for everything before it. The four job profiles at the fork (AI Engineer, GenAI Engineer, LLMOps Engineer, Agentic AI Engineer) are four emphases on these same nodes. Never compress these weeks to catch up elsewhere.
Week 26LLM 101
Speak the vocabulary of the role: training stages, tokens and cost, decoding, inference, memory.
9 lessons
Week 27Prompt engineering
Get reliable, structured, tool-using behaviour from models.
6 lessons
Week 28Context engineering + Project 4
Decide exactly what occupies the context window, and why.
5 lessonsBuilds: Schema-validated extraction CLI
Week 29Orchestration
Compose LLM calls, tools and data into reliable pipelines.
7 lessons
Week 30Vector databases
Choose embeddings, distance metrics and indexes with data, not vibes.
7 lessonsBuilds: Embedding model benchmark
Week 31RAG (1/2): building retrieval
Build naive RAG, watch it fail, and fix retrieval stage by stage.
6 lessonsNever cut
Week 32RAG (2/2) + Flagship 1
Advanced retrieval, grounded answers, and a production RAG system with an eval suite.
7 lessonsBuilds: Flagship 1: production RAG with evalsNever cutFlagship project
Week 33Memory, fine-tuning (light) and AI automation
Give systems memory, run one honest fine-tune, and build an automated workflow.
7 lessonsBuilds: QLoRA fine-tune with an honest eval
Week 34Agentic AI (1/2): planning, tools and state
Understand agents as loops, and engineer their tools like production APIs.
7 lessonsNever cut
Week 35Agentic AI (2/2): LangGraph, MCP + Flagship 2
Build a stateful agent with approval gates, tracing and a cost ceiling, plus an MCP server.
6 lessonsBuilds: Flagship 2: agent + MCP serverNever cutFlagship project
Week 36LLMOps (1/2): UX, deployment, observability and evals
Ship LLM features like production services, and prove changes with numbers.
7 lessonsBuilds: Evals wired into CI as a deploy gateNever cut
Week 37LLMOps (2/2): safety, governance and platforms
Defend LLM systems against attack and misuse, and run them responsibly.
7 lessonsBuilds: Optional: Spring AI RAG service (Java bridge)
8
Portfolio & Job Search
Weeks 38–40
Eight projects, two of them flagships. For each: an architecture diagram, stated tradeoffs, eval results and a cost analysis. The last two are what almost nobody includes. Start applying in week 39, not week 40: early interviews are diagnostic.
Week 38Project polish
Turn eight projects into evidence: diagrams, tradeoffs, eval results, cost analysis, demos.
5 lessons
Week 39Positioning and job prep
Reposition yourself as an AI engineer and start the interview drills.
6 lessonsBuilds: Résumé, GitHub and portfolio live
Week 40Interview and launch
Drill the classic questions, run mock interviews, and get applications out.
5 lessonsBuilds: Applications out

Roadmap structure adapted from the CampusX AI roadmap (AI Engineer fork).