Roadmap
AI Learning Path
From programming basics to building production LLM applications — a structured path through the concepts that matter.
Programming Basics
Variables, loops, functions, data structures, control flow.
Why
Every AI system is code. You need the foundation before touching models.
Build
Build a command-line number guessing game with user input and loops.
Python
Python syntax, OOP, list comprehensions, file I/O, virtual environments.
Why
Python is the dominant language for ML, data science, and AI engineering.
Build
Build a CSV data analyser that computes basic statistics.
NumPy
Vectors, matrices, broadcasting, dot products, element-wise ops.
Why
NumPy is the computational foundation for ML. Understand arrays before tensors.
Build
Implement matrix multiplication and softmax from scratch with NumPy.
Linear Algebra
Vectors, dot products, matrix multiplication, eigenvalues, SVD.
Why
Linear algebra is the mathematics of neural networks and embeddings.
Build
Implement a 2D transformation visualiser showing rotations and scaling.
Probability & Statistics
Probability, distributions, Bayes' theorem, entropy, KL divergence.
Why
LLMs are probabilistic systems. Loss functions, sampling, and logits all require this.
Build
Implement a Naive Bayes text classifier from scratch.
Machine Learning
Supervised/unsupervised learning, train/val/test split, overfitting, regularisation.
Why
Neural networks are a subset of ML. You need the broader context.
Build
Train a logistic regression model on a binary classification dataset.
Neural Networks
Perceptrons, dense layers, activations (ReLU, Sigmoid, Tanh), forward pass.
Why
The core building block of all modern AI systems.
Build
Build a 2-layer neural network from scratch using only NumPy.
Backpropagation
Chain rule, gradient computation, vanishing/exploding gradients.
Why
Backprop is how networks learn. Understanding it separates practitioners from users.
Build
Implement backpropagation manually for a 2-layer network, verify with numerical gradients.
Optimisation
SGD, momentum, Adam, learning rate schedules, batch size effects.
Why
Choosing the right optimizer and hyperparameters directly determines model quality.
Build
Compare SGD vs Adam on a noisy regression task. Plot loss curves.
Transformers
Attention mechanism, multi-head attention, positional encoding, encoder/decoder.
Why
Every major LLM — GPT, Claude, Gemini — is a Transformer variant.
Build
Implement scaled dot-product attention in pure Python/NumPy.
RAG
Embeddings, vector databases, semantic search, chunk selection, context injection.
Why
RAG is the dominant pattern for building LLM apps that work with private/recent data.
Build
Build a local Q&A system that retrieves from a document collection using embeddings.
AI Agents
Tool use, function calling, multi-step reasoning, agent loops, scratchpad.
Why
Agents are the next layer above RAG — AI that can take actions, not just answer questions.
Build
Build an agent that can use a calculator and web search tool to answer math questions.
LLM Product Engineering
Prompt engineering, streaming APIs, SSE, FastAPI, eval frameworks, cost management.
Why
Knowing the theory is not enough. Production AI products require full-stack engineering.
Build
Deploy a streaming RAG API with FastAPI and connect it to a Next.js frontend.