ROADMAP · ~30 hours over 6 weekends

Working SWE → AI engineer

Software engineers with 3-7 years shipping production code who are moving into AI features on their existing team.

Start: What is an LLM, really?17 stops · ~30 hours over 6 weekends

Who this is for

You’re a software engineer with 3-7 years shipping production code. You know git, code review, and how a deployed service degrades in ways the tests don’t catch. You’ve called an LLM API a few times, maybe built a chatbot, but you haven’t built an AI system that a team relies on.

The market wants you. Anecdotally: ~48% of AI-engineer job postings in 2026 target 3-7 YoE candidates — this is the biggest segment, larger than “entry-level” and “10+ YoE senior” combined.

What this roadmap does not try to be

Not “beginner AI.” If you don’t code, this isn’t the start. Try /roadmap/zero-to-ai.
Not “interview cram.” Two-week interview? Use /roadmap/interview-cram.
Not “research to production.” No paper reproductions, no ML theory beyond what production code needs.

Realistic timeline

About 30 hours of focused reading + hands-on. Distributed across 6 weekends at 5 hours each, you’re an AI-eng-capable engineer at the end. In practice, most people do this in 3-4 months of on-and-off Sunday afternoons, alongside their day job.

THE PATH

17 stops, in order

Phase 1Mental model2 stops

1
Lesson
What is an LLM, really?
Baseline mental model. Everything else assumes this.
2
Lesson
Reading a model card and a pricing sheet
How to skim a new model release in 60 seconds and know if it matters.

Phase 2Talking to LLMs4 stops

3
Lesson
Your first real API call
Reality check on how thin the SDK layer really is.
4
Lesson
Tokens, context, and what a call actually costs
You'll get asked "what's this going to cost?" on day one.
5
Lesson
Structured output you can trust
How to make an LLM output JSON your app can actually use.
6
Lesson
Streaming responses, honestly
Streaming isn't optional in 2026 UX; understand its cost + latency trade-offs.

Phase 3Tools & agents3 stops

7
Lesson
Tool use, from three lines of Python
The mechanism behind every agent, MCP server, and code executor.
8
Lesson
The agent loop, from scratch
The one architectural pattern that repeats everywhere.
9
Lesson
MCP, in the shape you'll use it
MCP is 2026's dominant tool integration protocol.

Phase 4Retrieval4 stops

10
Lesson
Embeddings, without the maths
If you skip this you'll ship a "just add RAG" solution that quietly fails.
11
Lesson
Vector stores worth using in 2026
The right vector store depends on scale + query pattern; know the choices.
12
Lesson
Chunking is the decision
80% of RAG failure modes trace to chunking.
13
Lesson
Evaluating a RAG system without lying to yourself
How to know your RAG works before your users tell you it doesn't.

Phase 5Production & proof4 stops

14
Lesson
Design an LLM gateway
The FLAGSHIP. Read this and you can design a production LLM system.
15
Interview card
How would you defend a customer-support agent against prompt injection when it can send emails and read a customer database?
The one prod-security question that comes up in every senior loop.
16
Case file
The eval pipeline that lied for a month
Read this before you build your own eval harness.
17
Lesson
The three portfolio projects that get you hired
Two projects that put you in the interview pool.

NOT COVERED HERE

What this roadmap skips

Training a model from scratch. This roadmap is engineering-around-models.
Prompt-engineering listicles. Where a technique matters it lives inside a lesson.
Deep RL / fine-tuning. Different arc; add if you specialize into it.

← All roadmaps