ai-hands-on (AI Engineering: Hands-on)
ramakm's Jupyter-notebook curriculum for AI engineering — "a complete, hands-on guide to becoming an AI Engineer" (MIT, 1.4k★ / 291 forks, 65 commits). Shared via Telegram, ingested 2026-08-22.
Its arc is six numbered folders worked in order — 1.Math (derivatives, vectors, gradients, matrix
operations, probability) → 2.PyTorch (tensors, reshaping, indexing) → 3.Neural-Network(NN) →
4.Transformer — then two run separately, 5.RAG and 6.OCR. A Start_here/learning_path.md gives the
recommended order, and a Basic ML Model Implementation folder adds linear and logistic regression,
decision trees and naive Bayes beside the deep-learning line.
Why it matters here: the first curriculum that is all three forms at once
The spoke's taxonomy has three forms and has treated them as alternatives: meta-curriculum contributes selection and order over other people's material; build-from-scratch contributes original lessons that make the reader implement the thing; executable-curriculum ships lessons that already run, where the reader's work is modification.
This repo is all three in one tree, and the split is legible folder by folder:
| Form | Where it shows up |
|---|---|
| Executable | The notebooks themselves, run in order under jupyter lab; contributing rules require "Restart & Run All before committing" |
| Build-from-scratch | 3.Neural-Network(NN) — "building neurons, layers, and networks from scratch" — and 4.Transformer, attention and multi-head attention by hand |
| Meta-curriculum | 19 external GEN AI reference links (Raschka, Alammar, Grootendorst, Karpathy's Build-GPT video, the HuggingFace and mlabonne courses), 6 recommended books marked "not included", and a 4-row ML-framework table |
So it inherits all three decay modes. executable-curriculum already records that each form rots
differently — a meta-curriculum by link rot, a from-scratch text when its toolchain moves, an executable
one when the API it calls changes underneath. A hybrid is exposed to every one of those at once: 19
outbound links that can break, PyTorch idioms that can shift, and a 5.RAG module calling live model
APIs. Nothing here is wrong with that; it is the maintenance bill, and it is roughly three times the bill
of a pure form. 65 commits is the budget being spent against it.
The scale contrast the spoke can now draw
ai-engineering-from-scratch (rohitg00) covers a near-identical arc — maths through LLM engineering — in 503 lessons across 20 phases with 17 capstones and ~33k stars. This one covers it in roughly ten notebook folders with 1.4k. Same claim of completeness, an order of magnitude apart in material.
That is not a quality judgment in either direction, and the spoke cannot make one: neither repo publishes completion data, and "complete guide" is a claim no source here supports with evidence. What the pair does show is that the word "complete" is doing no work in this genre — it describes intent, not coverage, and two curricula can both use it at 50× difference in size.
The gradient is steep, and unstated
1.Math covers derivatives, vectors and gradients; 4.Transformer builds decoder-only attention.
Between those sits one folder of PyTorch basics and one of neural-network construction. The README
addresses "a beginner or an engineer levelling up" and states no prerequisites at all. Compare
coding-interview-university and the roadmap sources, which are explicit about entry level, and
grinberg-graph-theory in ../mathematics-wiki, where a prerequisite list visibly decided the
syllabus. Here the entry bar is left to the reader to discover, which is the more common failure in
self-directed curricula and worth naming as such rather than treating as an omission unique to this repo.
The 1.Math folder is also the spoke's boundary with ../mathematics-wiki: it teaches calculus and
linear algebra as prerequisites to an engineering goal, where that spoke holds the same subjects
taught for themselves. The maths is a means here, which is why the page stays on this side.
Recorded, because a teaching repo's defaults are pedagogy
5.RAG defaults to Atlas Cloud (deepseek-ai/DeepSeek-V3-0324), with MiniMax (M3), OpenAI "or any
OpenAI-compatible API" as alternatives. The Atlas Cloud link in the README carries campaign-tracking
parameters naming this repo (utm_source=github&utm_medium=link&utm_campaign=ai-hands-on). Nothing states
a commercial arrangement and none is asserted here; what is observable is that the default provider in
a teaching module is reached through a tracked referral. Learners generally run the default, so a
curriculum's default is a pedagogical choice whoever benefits from it.
What it is not
No exercises with solutions, no assessment, no capstone, and no stated completion path — the reader finishes when the notebooks run out. Against ai-engineering-from-scratch's 17 capstones and the problem-archive form, this is a read-and-run curriculum: the notebooks are the whole contract. The author also maintains a separate book-reference repo and a Substack, so the reading list here is a pointer into a wider personal collection rather than a closed syllabus.
Related
executable-curriculum · build-from-scratch · meta-curriculum · ai-engineering-from-scratch · nirdiamant-prompt-engineering · problem-archive · ramakm