How the four terms nest inside each other
Picture four circles drawn one inside the next. The outer circle is artificial intelligence, which is any technique that lets a computer do something we would call intelligent, from a route planner to a chatbot. Inside it is machine learning, where the system is not handed the rules but learns them from examples. Inside that is deep learning, which does the learning with neural networks that have many layers. And inside deep learning sits generative AI, the models that create new text, images, audio or code.
In practice, an AI engineer is someone who builds products from these layers and picks the simplest one that solves the problem. Fraud scoring often needs classical machine learning on a table, reading a scanned form may need deep learning, and drafting a customer reply may need generative AI. A plain rule is sometimes the right answer.
There are two honest limits. First, the circles are a teaching picture, and real systems mix layers: one product may use a rule, a scikit-learn model and an LLM together. Second, LLMs, RAG and agents are not extra circles. They are things built with generative AI, and we explain them in their own section below.
The four layers side by side with one example each
Read across the cards and notice what changes: what the system is given, what it learns and what it produces.
Artificial intelligence, machines doing tasks that need judgement
The widest term. It includes hand-written rules, search and planning as well as learning systems. Example: a route planner that finds the shortest path between two Hyderabad addresses using a search algorithm, without learning anything.
Machine learning, learning patterns from examples
You give the algorithm examples and it finds the pattern. Typical work is predicting or classifying from tables. Example: estimating whether a delivery will be late from distance, weather and time of day. Tools: pandas and scikit-learn.
Deep learning, neural networks with many layers
A branch of machine learning that stacks layers of artificial neurons and needs plenty of data and computing power. It is strong at images, speech and language. Example: reading a number plate from a photograph. Tool: PyTorch.
Generative AI, models that create new content
Deep learning models trained to produce new text, images, audio or code, usually after learning from very large collections. Example: drafting a reply to a customer's email. Tools: LLM APIs and Hugging Face.
One problem solved six ways with restaurant reviews
Take one Hyderabad restaurant that collects hundreds of online reviews and wants to know what customers think. Here is how each layer would handle it.
- Rules, AI without learning: flag any review containing words such as cold, late or rude. It is cheap and easy to explain, but it misses sarcasm and every complaint it was not told about.
- Machine learning: label a few thousand reviews as positive or negative, then train a scikit-learn model that learns which words and patterns go with each label.
- Deep learning: train or fine-tune a neural network that reads whole sentences, so it can tell that not bad at all is praise.
- Generative AI: give an LLM the week's reviews and ask for a short summary of what customers liked and what went wrong, written in plain English.
- RAG: store the reviews as embeddings in a vector database. When the owner asks what customers say about parking, the system retrieves the relevant reviews first and the LLM answers from them, with quotes.
- Agent: a system that reads new reviews every morning, uses an LLM to decide which need a reply, drafts the reply, opens a ticket for serious complaints and asks the manager to approve before anything is posted.
Six questions that tell you which layer a product uses
Use these on any product, demo or job description. They turn buzzwords back into engineering.
Ask whether anything is learned from data
If the behaviour is fixed by hand-written rules, it is AI without machine learning. If the system was trained on examples, it is machine learning.
Ask what kind of data goes in
Rows and columns often mean classical machine learning. Images, audio and long text often mean deep learning.
Ask whether it creates something or picks a label
A label or a number is a predictive result. A new paragraph, image or piece of code is generative.
Ask whether it looks things up before it answers
If answers are drawn from your own documents, retrieved at the moment of the question, the system is using RAG.
Ask whether it acts in a loop
If it decides steps, calls tools and continues until a task is done, it is an agent, and it needs limits and approvals.
Ask how anyone knows it works
Test-set accuracy checks a machine learning model, test questions and human review check an LLM or RAG app, and logs plus limits keep an agent honest. This is where engineering begins.
LLMs, RAG and agents in plain words
A large language model, or LLM, is a neural network trained on very large amounts of text to predict the next token, a small piece of a word, over and over. That single skill produces fluent answers, summaries, translations and code. An LLM does not look facts up by default, so it can state wrong things confidently, which people call hallucination.
Retrieval-Augmented Generation, or RAG, reduces that problem. Documents are split into chunks and turned into embeddings, which are lists of numbers that capture meaning. When a question arrives, the system finds the closest chunks in a vector database such as pgvector and passes them to the LLM together with the question, so the answer is grounded in your material. RAG lowers hallucination but does not remove it, so you still test it with sample questions.
An agent puts an LLM inside a loop with tools. It reads a goal, chooses a step, calls a tool such as a database query or an API, reads the result and decides what to do next until the job is done. Frameworks such as LangGraph manage state and routing, and the Model Context Protocol connects agents to tools and data. Because agents can act, they need guardrails, retries, human approval for risky steps and defences against prompt injection.
Which layer to study first depends on who you are
You do not need to master all four at once. Start where your background gives you an edge.
Student who enjoys numbers and tables
Begin with classical machine learning using pandas and scikit-learn. Predicting from tables is a large part of everyday ML work and teaches evaluation properly.
Developer who wants to add AI features to an app
Focus on LLM APIs, embeddings and RAG, and keep your backend skills sharp. You will use deep learning models without training them from scratch.
Data or business analyst curious about AI
Learn machine learning basics and model evaluation, then try an LLM on the text data you already handle. Your domain knowledge is a real advantage.
Manager or founder who needs the vocabulary
Learn the four layers and the six questions above. When someone proposes an AI feature, ask which layer it uses and how success will be measured.
Tools you meet at each layer of the stack
Names change quickly, but these come up most in current AI engineering work.
- Rules and classic AI, plain Python logic and search algorithms
- Machine learning, pandas, NumPy and scikit-learn
- Deep learning, PyTorch, and TensorFlow in many teams
- Generative AI, LLM APIs and Hugging Face models
- RAG, embeddings, document chunking and PostgreSQL with pgvector or another vector database
- Agents, LangGraph, tool calling and the Model Context Protocol
- Running it all, FastAPI, Docker, MLflow, GitHub Actions and a cloud platform such as AWS or Azure
How the Madhapur AI and ML programme teaches each layer
Understanding the terms is a start. Our programme is designed so you build something at each level, as support and not a promise of any outcome.
Machine learning and deep learning basics in module three
Forty hours on feature engineering, supervised and unsupervised learning, evaluation, deep learning fundamentals with PyTorch, and model serving.
Generative AI, LLMs and RAG in module four
Forty hours on prompts, structured outputs, function calling, embeddings, vector databases, RAG, reranking and evaluation.
Agents in module five
Forty hours on tool calling, memory, routing, multi-agent patterns, human approval, guardrails and prompt injection defence, using LangGraph and MCP.
MLOps and LLMOps in module six
Thirty hours on experiment tracking, versioning, regression testing, CI/CD, tracing and cost control, so what you build keeps working.
Projects, mock interviews and placement support
You build a machine learning service, a RAG application and an agentic AI application, and mock interviews cover concept questions like this one. Placement support runs through a hiring-partner network as assistance only.
Quick answers about AI, ML, deep learning and generative AI
The short answers to the questions people type most when the terms get mixed up.
Is generative ai the same as machine learning?
No. Generative AI is a part of machine learning, usually deep learning, so it does use learning, but not all machine learning is generative. A model that predicts whether a loan will default generates nothing new, while an LLM that writes text does.
Is chatgpt machine learning or deep learning?
Chat assistants such as ChatGPT are built on large language models, which are deep neural networks trained with machine learning. So they belong to deep learning and to generative AI. The chat product also adds instructions, safety rules and sometimes tools around the model.
What is the difference between an llm and generative ai?
Generative AI is the wider category of models that create content such as text, images, audio or code. An LLM is one kind, a model specialised in language. An image generator is generative AI without being an LLM.
Is rag the same as fine tuning a model?
No. Fine-tuning changes a model by training it further on your examples. RAG leaves the model unchanged and supplies relevant documents at question time. RAG suits knowledge that changes often, such as policy documents, while fine-tuning suits teaching a style or task.
What is an ai agent and how is it different from a chatbot?
A basic chatbot answers a message and stops. An AI agent uses a model to plan, call tools, read the results and continue over several steps toward a goal. Because it can act, an agent needs limits, logging and human approval for risky steps.
Where to read next about the AI stack
Start with the programme page to see each layer as a module. The related guides cover the earlier basics, the engineer's role and the tools.
Build one small example of each layer this month
The clearest way to feel the difference is to build a tiny version of each: a rule, a scikit-learn model, an LLM call and a small RAG app. If you would like a guided path through all of them, talk to the admissions team about the AI and ML programme.