AI Engineer Roadmap

What Python is, why it is so popular, and how to install it and run your first script.

What Git is, installing it, the three areas, and making your first commit.

What machine learning is, and supervised, unsupervised and reinforcement learning explained with concrete examples.

The real end-to-end ML workflow: collecting and cleaning data, splitting, training, evaluating, deploying and monitoring.

What deep learning is, how it relates to machine learning and AI, why it took off, and where it is used today.

Code a two-layer neural network with only NumPy, train it on XOR, and see forward pass, loss and gradients in action.

Install PyTorch, create and reshape tensors, move them to the GPU, and understand why PyTorch is the most popular deep learning framework.

What generative AI is, how it differs from traditional AI, real-world use cases, and the skills students need to work with GenAI.

What a large language model is, how the transformer architecture works at a high level, and why scale matters.

What prompt engineering is, why it matters, and the anatomy of a good prompt: context, task, format and constraints.

Call an LLM API from Python, write a solid system prompt, stream the answer, and handle errors — a complete beginner project.

What Retrieval-Augmented Generation is, the problem it solves, and the query-retrieve-generate flow.

What embeddings are, cosine similarity, and how vector databases differ from normal databases.

What an AI agent is, how agentic AI differs from a chatbot or a fixed workflow, and when an agent is (and is not) the right tool.

Write a working AI agent in about 60 lines of Python — tools, the agent loop, stop reasons, errors and a step limit.

Create your first MCP server with the official Python SDK (MCPServer, formerly FastMCP), add tools with type hints, and test it with the MCP Inspector.

How low-rank adapters let you fine-tune billions of parameters on one GPU, and how QLoRA adds 4-bit quantization on top.

How LLMs are actually evaluated — perplexity, benchmark suites, human evaluation, and LLM-as-judge.

Latency and cost trade-offs, caching, batching, and quantization for running LLMs in production.

Serving a model online or in batch, versioning models and training data, and monitoring for drift.

Orchestrator–worker multi-agent designs, how to evaluate agents, and the guardrails every production agent needs.

Common agentic AI interview questions — tool calling, agent loops, memory, multi-agent design and safety — with clear answers.