Agent frameworks are getting easier to call and harder to understand, which may explain the sudden popularity of a from-scratch AI engineering curriculum on GitHub. The repository rohitg00/ai-engineering-from-scratch — slogan: "Learn it. Build it. Ship it for others." — picked up around 848 stars in a day.
The pitch is first principles. Rather than wiring together SDK calls, the curriculum walks learners through building AI systems by hand across four languages — Python, TypeScript, Rust and Julia. It is released free under the MIT licence, and community posts circulating alongside it describe roughly 523 lessons and an end-to-end, build-it-yourself approach.
That framing lands at an awkward but honest moment. The scarce skill in AI engineering is no longer calling a model; it is knowing what happens between the prompt and the answer — retrieval, memory, evaluation, batching, cost, failure modes. Engineers who can only drive a framework tend to discover what is missing when something breaks, or when the bill arrives. A curriculum that forces you to implement those pieces yourself is a reasonable antidote to framework fluency without comprehension.
The caveat is the usual one for fast-growing community projects: breadth this wide invites uneven quality, and a syllabus is not a verified credential. Treat it as a map of the territory and check each lesson against primary sources. It is trending for a reason, though — plenty of engineers want to know what is under the abstraction, not just how to call it.
