~/roadmap/llms $ self-learn --goal production-ready
Become a Production-Ready LLM Engineer
A complete self-learning course built from the 21-phase LLM roadmap — from Python and linear algebra all the way to fine-tuning, RAG, agents, and production deployment. Every phase tells you what to learn, why it matters, what to build, and when you're done. No fluff: each phase ends with something working in a repo.
00 How to work through this course
Three rules separate people who finish roadmaps like this from people who collect them. They come from how working LLM engineers actually learn on the job.
Every phase ends in a repo
Reading about attention is not knowing attention. Each phase has a Build task — finish it before moving on, even if it's ugly. Your GitHub history becomes your portfolio in Phase 21.
Don't skip the math or the transformer
Phases 2 and 6 are the two places learners cut corners — and the two places interviewers probe hardest. If you can derive backprop and write scaled dot-product attention from scratch, everything after gets easier.
Stuck for 2 days? Move on, mark it
Note the gap and keep going. Later phases revisit earlier ideas in applied form, and most "stuck" topics click the second time you meet them. Perfectionism, not difficulty, is the main dropout cause.
01 Pick your pace
The full roadmap is roughly 665 focused hours. Choose a weekly commitment and see what the finish line looks like. Consistency beats intensity — a pace you can hold for months wins.
Estimates assume focused, hands-on time (coding along, building the projects), not passive video watching. Add ~20% if you're also working full-time and studying tired — plan for it instead of quitting over it.
FINAL GOAL · PHASE 21 → GRADUATION
Build, fine-tune, evaluate, optimize & deploy — end to end.
Finishing the phases isn't the goal; being able to do the job is. You're done with this course when you can honestly say yes to each of these:
- I can train a small transformer from scratch and explain every tensor shape inside it.
- I can fine-tune an open model with LoRA/QLoRA on my own dataset — and prove it got better.
- I can build a RAG system with chunking, hybrid search, and reranking, and evaluate it with RAGAS.
- I can ship an agent that calls tools safely, with tracing, guardrails, and a cost budget.
- I can serve a model behind an API with batching, quantization, and latency/cost numbers I measured myself.
- I have a public portfolio of these projects, written up so another engineer can reproduce them.