How I'd become an AI engineer in 2027 (all free)

Read a few AI engineer job posts in Pakistan and you’ll notice something. The required section is mostly a backend job: Python, FastAPI, a SQL database, Redis, Docker, tests, Git. The AI skills (LLMs, RAG, agents) usually sit under preferred.

So most students learn this backwards. They start with machine learning theory, and the job wants a backend engineer who can build with AI. This roadmap has one rule: backend first, AI on top.

I’m a backend engineer at a software house in Lahore, and this is the plan I’m following myself. Five stages, about twelve months, one project and one checkpoint per stage. Don’t move to the next stage until you pass the checkpoint.

First, what an AI engineer actually is

• AI researcher: invents new models. Usually a PhD. Not this roadmap.
• ML engineer: trains custom models on data (fraud detection, recommendations).
• Data scientist: analyses data to answer business questions.
• AI engineer: takes models someone else trained (GPT, Claude, Gemini, open-source models) and builds real products on top of them: assistants that answer from documents, agents that finish tasks.

If you’re doing BS AI and think you must train models from scratch to get hired: for AI engineer roles, you mostly don’t. You need to build reliable software around models.

Stage 1: Python and backend (months 1 to 3)

Learn Python properly: functions, classes, error handling, files and JSON, virtual environments. Then FastAPI, Postgres and Git from day one.

• CS50’s Introduction to Programming with Python (Harvard, free)
• FastAPI official tutorial (one of the best free tutorials on the internet)

Project: a search API over real data. Collect your university’s course list or past-paper titles, store them in Postgres, build endpoints to search them, push it to GitHub.

Checkpoint: you can build an API with a database without copying a tutorial line by line.

Already know Rails, .NET or Node? You’re not starting from zero. The concepts carry over. Add Python and FastAPI on top; don’t throw away what you know.

Stage 2: Working with LLMs (months 4 to 5)

Learn tokens, context windows, temperature, why a model sounds confident when it’s wrong, and how you pay per token. Then prompting properly: clear instructions, examples, and structured output (asking for JSON your code can use).

No dollar card? Two fixes:

• The Gemini API has a free tier, and Pakistan is on its supported regions list.
• Ollama runs small open models on your own laptop, free.

You don’t need a GPU to learn this. Most of the job is calling models, not training them.

• Andrej Karpathy on YouTube for intuition on how LLMs work
• DeepLearning.AI short courses for the practical side (check which ones are free when you enrol)
• Hugging Face LLM Course (free)

Project: paste a messy job post in, get clean JSON back: role, skills, experience, salary if listed.

Checkpoint: you can explain tokens and context windows to a friend, and your app gets valid JSON from the model every time.

Stage 3: RAG and evals (months 6 to 8)

This is the stage that gets you hired.

RAG (retrieval augmented generation): the model doesn’t know your data, so you search your documents first, find the relevant chunks, and give those to the model with the question. Learn embeddings, chunking and a vector database. Start with pgvector: it lives inside the Postgres you learned in Stage 1.

Evals: write 20 questions where you already know the right answer. Every time you change the prompt, the model or the chunking, run all 20 and record the score. This is the difference between a demo and a product, and almost nobody does it.

Your unfair advantage: Urdu and Roman Urdu. Most tutorials only build for English. A RAG app that answers from your university handbook in Roman Urdu is something a Pakistani company understands instantly, and most applicants can’t build it.

Project: that Roman Urdu handbook assistant, with your eval score in the README.

Checkpoint: a working RAG app and a measured score on your 20 questions.

Stage 4: Agents and MCP (months 9 to 10)

An agent is a model that can use tools (search the web, query a database, call an API) and keeps going until the task is done.

• Learn tool calling by hand first, before any framework, so you understand what’s happening.
• Then one framework: LangGraph.
• Learn MCP (Model Context Protocol), the standard way to plug tools and data into AI apps.
• Hugging Face AI Agents Course (free)

Project: a job-hunt agent. Every morning it checks a few job boards, filters posts by your skills, and sends you the best five. Reuse your Stage 2 project to read each post.

Checkpoint: the agent finishes a multi-step task on its own, and you can explain every tool call it made.

Stage 5: Ship it and get seen (months 11 to 12)

Production means it works when you’re not watching:

• Put it in Docker (get started).
• Deploy somewhere free, like Hugging Face Spaces (the basic CPU tier is free).
• Keep API keys out of your code, add logging, and handle the model being slow or down.

All year, in the background:

• Problem solving: most software house interviews still have a problem-solving round. Three or four problems a week is enough. Don’t leave it for the last month.
• Learn in public: post every project on GitHub and LinkedIn as you finish it. Five shipped projects and a story beat a CV full of course names.

The one project I’d put on my CV

One app that uses all five stages: a Roman Urdu study assistant for your university.

1. FastAPI + Postgres backend (Stage 1)
2. Structured outputs (Stage 2)
3. RAG over the handbook, past papers and course outlines, with an eval score (Stage 3)
4. An agent that can also check deadlines or the timetable (Stage 4)
5. Deployed, with Docker and logging, and real students using it (Stage 5)

In the interview you don’t say “I did a course on RAG.” You say “I built this, students used it, and my score went up after I fixed my chunking.” You earn those numbers yourself, and that sentence gets you hired.

Two mistakes that waste your first six months

• Collecting certificates instead of building. Job posts ask what you built, not which course you finished.
• Vibe coding everything. Use Claude Code or Cursor, but if you can’t explain the code, you can’t debug it, and the interview will expose that in five minutes. AI does the typing. You do the thinking.

Links and free tiers change. Check each one is still live and free on the day you start.