AI Engineering from Scratch
rohitg00's hands-on AI/ML engineering curriculum — 503 lessons across 20 phases, ~33k stars.
Parked in the hub _inbox on 2026-06-15 under learning-roadmaps; ingested here on the 2026-08-04
spin-out.
The arc
Mathematics (linear algebra, calculus, probability) → classical ML → deep learning (neural networks from scratch, then PyTorch/JAX) → computer vision, NLP, speech, RL → LLM engineering (build-from-scratch, tokenization, fine-tuning, RLHF, quantization) → production (RAG, agents, MCP servers, infrastructure) → safety and alignment. 17 capstones.
Why it is the cluster's outlier
The other two founding sources are meta-curricula pointing at other people's material and shaped by an exam. This one is neither.
Its pedagogy is stated and distinctive: build-it → use-it → ship-it — implement the thing by hand, then use the framework equivalent, then produce a reusable artifact. "You don't just learn AI, you build it by hand." The output is a portfolio, not readiness to be questioned.
And there is no standardized AI-engineering interview to prepare for. The absence of an assessment regime and the presence of construction-based pedagogy arrive together here, which is the observation the spoke's synthesis is built on — and, on three sources, an observation rather than a finding.
Claim re-checked 2026-09-04 and it still holds. A search for an AI-engineering assessment standard returns company-specific rubrics (Meta's rounds, individual firms' scoring sheets) and recruiter content, plus one recurring format recommendation — work sample, structured behavioural, pair programming, system design. That is advice about interviewing, not a shared exam anyone prepares against, and it is exactly what the LeetCode canon is and this subject lacks. Nothing to retract; the date moves.
Where it stops being this spoke's business
The subject matter belongs to ../machine-learning-wiki, which owns machine learning as a discipline.
This page is about the curriculum — its sequence, its pedagogy, its scope. A source teaching a
technique to make a claim about that technique routes there instead.
Tier
T1, re-tiered 2026-08-21 under tier by the claim, not the URL. The load-bearing facts are the curriculum's inventory and design — 503 lessons across 20 phases, 17 capstones, the arc from linear algebra to alignment, and the stated build-it → use-it → ship-it rule. Those are the artifact describing itself and they are checkable in the repository, which is what the rule puts at T1.
The T3 half is still here and still flagged: "You don't just learn AI, you build it by hand" is an effectiveness claim for a curriculum, and there is no evidence anyone completes 503 lessons. The spoke's thesis about popularity measuring intent rather than learning (meta-curriculum) is the wiki's own finding about this source class. It belongs in that prose, not in this page's grade.
Related
meta-curriculum · coding-interview-university · awesome-system-design-resources