Operating Plan · Learn AI project
A 26-week execution plan for going from beginner coder to someone who ships real AI products. Not a syllabus to admire — a contract with yourself.
Here's the honest version. You are a beginner coder who wants to build and ship AI products, with real time to put in — 20-plus hours a week is close to a part-time job. That is enough to get genuinely good in six months, but only if you spend it building, not collecting courses. The single most common failure mode at your stage isn't lack of talent, it's tutorial hell: watching lecture after lecture, feeling productive, and having nothing to show for three months except notes. This plan is built to make that failure mode structurally hard to fall into — every phase below ends with something running, deployed, or public, not just "I understand it now."
You have a decent math foundation already, so this plan doesn't stop to re-teach you calculus from zero. It picks up the linear algebra and probability you need exactly when a phase needs it, and no earlier — math for builders is a tool you sharpen on the way to the job, not a gate you have to clear first. The same discipline applies everywhere: depth is earned by what the current project demands, not by what looks impressive on a roadmap.
Six months of consistent building will put you meaningfully ahead of almost everyone who "learns AI" by consuming content. It will not make you a research scientist, and it doesn't need to — your goal is products, not papers. Hold yourself to the checklist, not the vibe of feeling like you're learning.
Python fluency and real developer tooling: git & GitHub, the terminal, virtual environments, an editor you're fast in. Enough programming fundamentals — functions, data structures, file I/O, error handling — that code stops feeling like magic.
REST APIs, JSON, the `requests` library, pandas basics for wrangling data, and just enough front-end (Streamlit or Gradio, not raw HTML/CSS/JS) to put a UI on top of your code. Then: your first call to an LLM API and your first deployed AI app.
Linear algebra and probability, but only the parts ML actually uses: vectors, matrices, dot products, gradients, distributions, Bayes' theorem — applied immediately, not proved in the abstract. Then classical ML with scikit-learn: regression, classification, clustering, and the discipline of proper train/validation/test evaluation.
Neural networks from the ground up in PyTorch, then transformers, attention, and tokenization — enough that "how does an LLM work" stops being a black box. The goal isn't to train GPT from scratch, it's to never again treat a model like magic when you're debugging your own product.
This is the phase that actually matters for your goal. Rigorous prompt engineering, retrieval-augmented generation (RAG) and vector databases, tool/function calling, agents that take multi-step actions, evaluating LLM outputs properly instead of vibes-checking, and a first fine-tune.
Everything before this was preparation. Pick one product idea, constrain the decision to a week, and build an MVP using what you now know. Deploy it, get it in front of strangers, and iterate on real feedback — not more building in isolation.
Open-ended by design. What you build here depends on what Product #1 actually revealed you're missing — that's more honest than guessing now. Likely candidates: production-grade evals and monitoring, cost and latency optimization, fine-tuning open models yourself, multi-agent systems, voice or vision. The loop from here is simple: build, ship, learn, repeat — and revisit this plan monthly to be honest about where you actually are.
If you fall behind — and at some point you will — the fix is never to skip a project to catch up on the calendar. Cut scope inside the current project until it fits the time you actually have, ship the smaller version, and move on. A shrunk deliverable that shipped beats a full one that's still "almost done" in your head. The one thing you're not allowed to cut is shipping itself.
Twenty-six weeks assumes real, defended hours — not hours that evaporate into browsing AI Twitter and calling it research. Protect the time the same way you'd protect a job you actually wanted.