Vinay C.

Start here · Free

The GenAI Engineer Roadmap.

If you write software and want to move into AI engineering, the problem isn't a shortage of material — it's that none of it is in order. This is the order. Five stages, 25 pieces I've written from production work, and one thing to build at each stage so you finish with evidence instead of tab history.

Stuck? Message me on WhatsApp

Before you start

  • You do not need an ML degree. Almost all of this is systems work — APIs, data pipelines, evaluation, latency and cost. Your backend experience is the advantage, not the gap.
  • Go in order. Every stage assumes the one before it. Stage 04 is where most people start and it's why they stay confused.
  • Build the thing at each stage. Nobody hires on reading. The build is the part that shows up in your interview answers.
  • Expect three to six months at a few hours a week. Anyone promising you six weeks is selling something.

The five stages

01 — 05

01

Foundations

Most people skip straight to frameworks and spend the next six months confused by their own bugs. Learn what the model is actually doing first.

Read, in this order

  1. Large Language Models (LLMs) – Introduction3 min
  2. Understanding Temperature, Top-p, and Top-k in Large Language Models3 min
  3. 3 Common LLM Myths which was surprising!4 min
  4. 10 LLM application patterns you must know1 min

You'll be able to

You can explain tokens, context windows, and sampling to someone else without hand-waving — and you know which problems an LLM is the wrong tool for.

Build this

Call an LLM API directly, no framework. Change temperature and top-p, and watch what happens to the same prompt ten times.

02

Embeddings & retrieval

Nearly every "the AI is hallucinating" bug in production is a retrieval bug. This stage is where the money is, and it is the one people rush.

Read, in this order

  1. Your LLM Has No Memory. Embeddings Are the Fix.4 min
  2. My 6 Learnings About Embedding Models (and Why Many RAG Systems Struggle)3 min
  3. FAISS - What AI Engineer needs to know!6 min
  4. Fast or accurate? Why does good retrieval use two models3 min

You'll be able to

You can pick an embedding model on evidence rather than a leaderboard screenshot, and explain why a bi-encoder and a reranker do different jobs.

Build this

Index 200 documents you care about. Search them. Find ten queries that return the wrong thing, and work out why for each one.

03

RAG in production

A RAG demo takes an afternoon. A RAG system you can change without breaking is a different discipline — chunking, parsing, and above all evaluation.

Read, in this order

  1. RAG Pipeline Explained: How Retrieval-Augmented Generation Makes AI Answers More Useful6 min
  2. Building Production RAG Systems2 min
  3. Best PDF Parser for Better LLM and RAG3 min
  4. OKF and RAG: What I Learned While Thinking Through Agentic Knowledge Systems9 min

You'll be able to

You can say how well your system retrieves, with a number, and prove a change made it better. This is the single skill that separates candidates.

Build this

Write 50 real questions with known-correct answers. Score your pipeline on them. Change one thing. Score it again.

04

Agents & frameworks

Now the frameworks make sense, because you know what they are hiding. Start with the Honest Guide series in order — it exists so you can skip the dead ends.

Read, in this order

  1. Honest Guide - LangChain for Beginners20 min
  2. Honest Guide - Langchain for Intermediate25 min
  3. Honest Guide - LangGraph for Beginners17 min
  4. Honest Guide — Google Agent Development Kit (ADK) for Beginners18 min
  5. 10 AI Agent Design Patterns Every AI Engineer Should Know3 min
  6. MCP Architecture2 min
  7. What is an AI harness, and why does it matter more than your model choice4 min

You'll be able to

You can design an agent loop with real tools, state, and a human in it, and argue for or against a framework instead of defaulting to one.

Build this

One agent that does something genuinely useful to you every week. Ship it, then instrument it so you can see when it fails.

05

Systems & the interview

AI engineering interviews are still systems interviews with an AI layer on top. Cost, latency, and scale questions are where strong candidates separate.

Read, in this order

  1. Zero Token Architecture: AI engineers should always start with this!2 min
  2. How I Would Design Google Drive at Scale (A System Design Case Study)6 min
  3. System Design Interview: Designing Seat Booking for BookMyShow6 min
  4. IRCTC’s redesign: what may have changed behind the scenes to handle 10x scale6 min
  5. Demystifying AWS Messaging: SNS vs. SQS2 min
  6. Bloom Filters: A Space-Efficient Data Structure2 min

You'll be able to

You can whiteboard an AI feature end to end — retrieval, caching, cost per request, failure modes — and defend the trade-offs.

Build this

Design one system a week out loud, on paper, timed at 45 minutes. Record yourself. Watching it back is unpleasant and it works.

Worth reading, off the critical path

All notes →

When reading isn't enough

Want someone to check your work?

The roadmap is free and always will be. If you want your resume torn apart, a mock interview with someone who runs the real loop, or a monthly 1:1 to keep you honest — that's what mentoring is for.

See mentoring →