---
title: "265 Topics to Master AI Engineering (and the 6-Week Order I'd Learn Them In)"
description: "265 AI engineering topics from roadmap.sh's AI Engineer (189) and AI Agents (100) roadmaps, deduped, plus the 6-week order to learn them and what to build each week."
date: "2026-11-15"
type: "resource"
tags: ["ai engineering", "roadmap", "llms", "rag", "mcp", "agents", "evals"]
legacy: true
---

      <p>
        You commented <strong>topics</strong>, so here's the full map. roadmap.sh has two free
        roadmaps that together cover what an AI engineer actually does in 2026: the
        <a href="https://roadmap.sh/ai-engineer">AI Engineer roadmap</a> (189 topics) and the
        <a href="https://roadmap.sh/ai-agents">AI Agents roadmap</a> (100 topics). I counted every
        topic in the source repo on 15 November 2026 and merged the ones that appear in both or are
        the same topic under a different name. That leaves <strong>265 distinct topics</strong>, with
        660 learning links between them.
      </p>
      <p>
        Click any box on those two roadmaps and you get a short intro plus free links for that topic.
        This post is not a copy of the roadmaps (their content stays on roadmap.sh). It's the order I'd
        go through them in, and what I'd build each week so it sticks.
      </p>

      <div class="highlight">
        <p style="margin-bottom: 0">
          <strong>The short version:</strong> don't try to tick off 265 boxes. Learn how LLMs work,
          then context, retrieval, tools, evals and security, in that order. Build one small thing
          every week. By week 6 you have a working agent on your GitHub, and that's what gets you
          the interview.
        </p>
      </div>

      <h2 id="roadmaps">The two roadmaps</h2>
      <h3>roadmap.sh: AI Engineer <span class="stars">189 topics</span></h3>
      <p class="repo-link"><a href="https://roadmap.sh/ai-engineer">roadmap.sh/ai-engineer</a></p>
      <p>From how LLMs work and picking a model, to prompting, context engineering, embeddings and RAG, MCP, evals, monitoring, safety and multimodal apps.</p>
      <h3>roadmap.sh: AI Agents <span class="stars">100 topics</span></h3>
      <p class="repo-link"><a href="https://roadmap.sh/ai-agents">roadmap.sh/ai-agents</a></p>
      <p>The agent side: the agent loop, tools, memory, MCP, agent frameworks, testing and the security of letting a model act for you.</p>
      <h3>nilbuild/developer-roadmap <span class="stars">~372K ★</span></h3>
      <p class="repo-link"><a href="https://github.com/nilbuild/developer-roadmap">github.com/nilbuild/developer-roadmap</a></p>
      <p>The open repo behind roadmap.sh (it used to live at kamranahmedse/developer-roadmap). Every topic is a file here, which is how I counted them.</p>

      <h2 id="start">The 4 I'd start with</h2>
      <ol>
        <li><strong>Context engineering.</strong> What goes into the context window, how to compact it when it fills up, how to keep agents from polluting each other's context, and the ways it fails. Most "the AI is dumb" bugs are context bugs.</li>
        <li><strong>MCP.</strong> Build one MCP server and one client yourself. Once you've done both, connecting a model to any tool or database stops being magic. Start at <a href="https://modelcontextprotocol.io/">modelcontextprotocol.io</a>.</li>
        <li><strong>Evals.</strong> Deterministic checks, model-graded checks and human review. If you can show how you measured your AI feature, you're ahead of most candidates. <a href="https://github.com/confident-ai/deepeval">DeepEval</a> (~19K ★, Apache-2.0) is an easy place to start.</li>
        <li><strong>Agent memory and security.</strong> Short-term vs long-term memory, and then prompt injection and tool sandboxing. An agent that can act needs limits on what it can touch.</li>
      </ol>

      <h2 id="plan">The 6-week order</h2>
      <h3>Week 1: How LLMs actually work</h3>
      <p>Tokens, context windows, sampling (temperature, top-p), open vs closed models, and how pricing works per token.</p>
      <p><strong>Build:</strong> a CLI that sends the same prompt to two models and prints the answer, the tokens used and the cost.</p>
      <h3>Week 2: Prompting, then context engineering</h3>
      <p>Few-shot and chain-of-thought prompting, structured outputs, then the bigger idea: managing everything the model sees, not just the prompt.</p>
      <p><strong>Build:</strong> a summariser for your college notes that returns clean JSON every time.</p>
      <h3>Week 3: Embeddings, vector databases and RAG</h3>
      <p>Embeddings, chunking, similarity search, a vector database, and when RAG beats fine-tuning.</p>
      <p><strong>Build:</strong> "chat with my syllabus": upload PDFs, ask questions, show the source for every answer.</p>
      <h3>Week 4: Tools, MCP and the agent loop</h3>
      <p>Function calling, the reason, act, observe loop, and MCP servers and clients.</p>
      <p><strong>Build:</strong> an MCP server for something you use daily (your calendar, a GitHub repo, a Notion page) and an agent that uses it.</p>
      <h3>Week 5: Evals and observability</h3>
      <p>Test sets, model-graded evals, tracing every call, and tracking cost and latency.</p>
      <p><strong>Build:</strong> 30 test questions for your week 3 app, a score you can re-run after every change, and traces in <a href="https://github.com/langfuse/langfuse">Langfuse</a> (~37K ★).</p>
      <h3>Week 6: Security, cost and shipping</h3>
      <p>Prompt injection, permissions for tools, rate limits, and keeping API keys out of the browser and out of git.</p>
      <p><strong>Build:</strong> deploy the agent from week 4 with a live link, a README, and one paragraph on how you tested it.</p>
      <p>You don't need all 265. You need one deployed agent and the numbers to show it works.</p>

      <h2>How I'd use this</h2>
      <ul>
        <li>Open the two roadmaps side by side and use this order to decide what to click next.</li>
        <li>Mark a topic done only when you've used it in code, not when you've watched a video on it.</li>
        <li>Want to run models locally while you learn? <a href="https://github.com/ggml-org/llama.cpp">llama.cpp</a> (~135K ★, MIT) just hit v1.0.</li>
        <li>Post each week's build in public. Six posts, six repos, one story for your interviews.</li>
      </ul>

      <p>
        Next: <a href="/blog/ai-engineering-from-scratch-free-course">523 free lessons to become an AI engineer</a>.
      </p>

      <p>
        Follow
        <a href="https://www.instagram.com/avi_vashishta29/">@avi_vashishta29</a>
        for one practical AI dev resource a day.
      </p>
