AI & Machine Learning
From zero to production-grade AI/ML — 60+ modules, no prerequisites
60+ Modules. Zero to Advanced.
Follow in order. Each module builds on the last. Every concept is introduced exactly when you need it, not before.
What is AI? ML, DL and GenAI Explained
The hierarchy explained from scratch — what AI, ML, Deep Learning, and Generative AI actually mean, how they relate, and why everyone gets them confused.
The AI/ML Landscape — Tools, Roles, and Career Paths
Every tool, every role, every career path in the AI/ML ecosystem mapped clearly. Know exactly where you fit before you write a single line of code.
Vectors, Matrices and Tensors
The language every ML algorithm speaks. From a single number to multi-dimensional arrays — with visual intuition first, formula second.
Matrix Multiplication and Linear Transformations
How data flows through neural network layers. One of the most important operations in all of ML — explained visually before mathematically.
Dot Product and Similarity
The operation behind every recommendation engine, embedding search, and attention mechanism. Cosine similarity, projections, and geometric intuition.
Eigenvalues and Eigenvectors
The mathematical foundation of PCA, spectral clustering, and page rank. What eigenvectors are, why they matter, and how to compute them.
Derivatives, Gradients and the Chain Rule
The mathematical engine behind every learning algorithm. Understand gradient descent before you ever run model.fit().
Probability Distributions and Bayes Theorem
How ML models reason under uncertainty. Normal distribution, Bernoulli, Poisson, and why Bayes theorem appears in almost every algorithm.
Information Theory — Entropy, Cross-Entropy and KL Divergence
Why loss functions in neural networks look the way they do. The information theory behind cross-entropy loss explained from scratch.
Python for Machine Learning
Not Python 101. Python for ML — NumPy vectorisation, Pandas DataFrames, matplotlib visualisation, and the sklearn interface that every algorithm shares.
NumPy Arrays and Broadcasting
The backbone of all numerical ML. Arrays, indexing, slicing, broadcasting, and vectorised operations that replace for loops at scale.
Pandas for Data Analysis
Load, clean, transform and explore datasets. GroupBy, merge, pivot, apply — the operations every ML project starts with.
Data Visualisation for ML
Plot distributions, correlations, confusion matrices, ROC curves, and learning curves. See what is happening in your data before you model it.
Scikit-learn Interface
The API every sklearn algorithm shares. fit, transform, predict, Pipeline, ColumnTransformer — understand the interface once and every algorithm becomes obvious.
Data Collection — APIs, SQL, Files and Scraping
Where ML data actually comes from. Pull from REST APIs, query databases, read Parquet files, and scrape web data — all with production-grade Python.
Missing Values, Outliers and Data Cleaning
The step that determines whether your model works or fails silently. Every imputation strategy, when to use each, and what breaks when you skip this.
Feature Scaling — Standardisation and Normalisation
Why scale matters, what StandardScaler and MinMaxScaler actually do, and which algorithms break without scaling.
Encoding Categorical Features
One-hot encoding, ordinal encoding, target encoding — what each one does to your data and when each one is the right choice.
Feature Engineering and the Sklearn Pipeline
Create new features, combine transformers, and build a single reusable pipeline that preprocesses and models together.
Train / Validation / Test Split
Why three splits, not two. Holdout sets, stratified splits, data leakage across splits, and the time-series exception where random splits break everything.
What is Machine Learning?
Not the Wikipedia definition. The actual idea — what training means mechanically, the 3 types of ML, the 7-step workflow, and 12 key terms defined once and for all.
Linear Regression — Swiggy Delivery Time Prediction
The first ML algorithm. Build a working Swiggy delivery time predictor from scratch — full intuition, math, code, 6 real errors, and a 5-day job walkthrough.
Logistic Regression — Razorpay Fraud Detection
When the output is a category, not a number. Sigmoid function, decision boundaries, and why logistic regression is still the first thing you try on classification problems.
Decision Trees — Loan Approval at HDFC
The algorithm that thinks in if-then questions. Gini impurity, information gain, pruning, and why decision trees are the foundation of every ensemble method.
Support Vector Machines — Margin Maximisation
The algorithm that finds the widest possible boundary between classes. Kernels, the kernel trick, and why SVMs still matter for small high-dimensional datasets.
K-Nearest Neighbours — Similarity-Based Prediction
The simplest possible ML algorithm — predict based on what your neighbours look like. Distance metrics, the curse of dimensionality, and when KNN actually works.
Naive Bayes — Probabilistic Text Classification
Bayes theorem applied to classification. Why the naive independence assumption works surprisingly well for spam filters and document classification.
Random Forest — Zepto Stock Prediction
Bagging + random features = a model that beats most others on tabular data. Feature importance, out-of-bag evaluation, and hyperparameter intuition.
Gradient Boosting — How XGBoost and LightGBM Work
The most powerful classical ML algorithm family. Sequential weak learners, residuals, and why gradient boosting wins almost every tabular ML competition.
XGBoost in Practice — End to End
Train, tune, and interpret XGBoost on a real dataset. Regularisation parameters, early stopping, SHAP values, and production deployment.
LightGBM — Fast Gradient Boosting at Scale
Microsoft's answer to XGBoost. Leaf-wise tree growth, histogram-based splitting, and why LightGBM trains 10x faster on large datasets.
K-Means Clustering — Customer Segmentation
Finding hidden groups in data without labels. Inertia, elbow method, silhouette scores, and when clustering is and isn't the right approach.
PCA — Dimensionality Reduction
Turn 100 features into 10 without losing most of the information. Explained variance, scree plots, and when PCA helps and when it hurts.
Evaluation Metrics — Beyond Accuracy
What accuracy hides, and what every other metric reveals. The confusion matrix explained from scratch, with Razorpay fraud detection as the running example.
Calibration — Are Your Probabilities Trustworthy?
Reliability diagrams, Brier score, and Platt scaling vs isotonic regression — when your model says 80% fraud probability, does it actually mean 80%?
ROC Curve and AUC — Threshold-Independent Evaluation
Why ROC-AUC is the most trustworthy single number for a classifier. How the curve is constructed, what the area means, and when to use PR-AUC instead.
Cross-Validation and the Bias-Variance Tradeoff
Why train/test split alone is not enough. K-fold, stratified K-fold, time-series split, and the bias-variance tradeoff visualised clearly.
Hyperparameter Tuning with Optuna
From grid search to Bayesian optimisation. How Optuna works, when to use it, and how to tune any model without wasting GPU hours.
Model Interpretability — SHAP and LIME
Why your model makes the predictions it does. SHAP values for any model, LIME for local explanations, and how to present findings to non-technical stakeholders.
Regression Metrics — MAE, RMSE, R²
How to measure how wrong a regression model is. MAE, RMSE, and R² explained with intuition, formulas, and when to use which.
Neural Networks from Scratch
Build a neural network from nothing but NumPy. Layers, weights, activations, forward pass — no black boxes, every line of code explained.
Backpropagation — How Neural Networks Learn
The chain rule applied to a network of layers. Gradients flow backward, weights update, the network gets better. Understood once, never forgotten.
Activation Functions and Loss Functions
Why ReLU replaced sigmoid, what GELU is doing in Transformers, and which loss function to use for which problem. The full reference.
Optimisers — SGD, Adam, AdamW
How neural networks update their weights. Momentum, adaptive learning rates, weight decay — and why Adam is the default for almost everything.
Batch Normalisation and Dropout
Two techniques that made deep networks trainable. What batch norm actually normalises, why dropout works as an ensemble method, and when to use each.
CNNs — Meesho Product Image Classification
Convolutions, filters, pooling, feature maps, and the ResNet skip connection that let neural networks go from 5 layers to 1000. Full PyTorch code.
RNNs and LSTMs — Sequence Modelling
When order matters. The vanishing gradient problem, LSTM gates (input/forget/output/cell state), and when RNNs are still the right choice over Transformers.
Transformers and Self-Attention
The architecture behind GPT, BERT, and every modern LLM. Query-Key-Value attention, multi-head attention, positional encoding — built up from scratch.
Tokenisation and Word Embeddings
How text becomes numbers. BPE, WordPiece, SentencePiece tokenisers explained. Word2Vec, GloVe, and contextual embeddings — why each one exists.
BERT and the Encoder-Only Family
Masked language modelling, next sentence prediction, fine-tuning on downstream tasks. The model that changed NLP — still powering classification and NER.
Fine-Tuning with PEFT — LoRA and Adapters
Tune less than 1% of a model's parameters and get 95% of the performance. LoRA, adapters, and prefix tuning — when and how to use each.
RAG — Retrieval-Augmented Generation
Give any LLM external memory. Vector databases, semantic search, chunking strategies, and the full RAG pipeline from document to answer.
Prompt Engineering
Zero-shot, few-shot, chain-of-thought, ReAct — the patterns that consistently improve LLM outputs. With real before/after examples for every technique.
LLM Agents and Tool Use
When LLMs call functions, search the web, write and run code, and plan multi-step tasks. The architecture behind every AI agent.
Image Fundamentals — Pixels, Channels and Tensors
How computers see images. Pixel values, colour channels, image tensors, normalisation, and the preprocessing pipeline every vision model expects.
Data Augmentation — Training on Limited Image Data
How to multiply your image dataset without collecting more data. Flips, crops, colour jitter, mixup, cutout — and how each one affects what the model learns.
Object Detection — YOLO and Feature Pyramids
Detect and localise multiple objects in one pass. Anchor boxes, IoU, non-maximum suppression, and why YOLO became the production standard.
Semantic Segmentation — Pixel-Level Classification
Label every pixel in an image. U-Net architecture, skip connections, and how segmentation powers medical imaging and autonomous vehicles.
Transfer Learning — Fine-Tuning Pretrained Vision Models
Get ImageNet-level features without ImageNet-level compute. Feature extraction vs fine-tuning, layer freezing, and choosing the right backbone.
What is Generative AI?
The shift from judging to creating. GANs, VAEs, diffusion, and LLMs — what makes each one generative, and when each one is the right architecture.
GANs — Generator vs Discriminator
Two networks in adversarial competition. Mode collapse, training instability, Wasserstein distance — the honest account of what makes GANs hard to train.
Variational Autoencoders — Learning Latent Representations
Encode data into a structured latent space and decode it back. The reparameterisation trick, KL divergence loss, and why VAEs enable controllable generation.
Diffusion Models and Stable Diffusion
Forward noise, reverse denoising, DDPM, DDIM, latent diffusion, classifier-free guidance. How Stable Diffusion generates photorealistic images from text.
LLMs — Pretraining, RLHF, and Scaling Laws
How GPT, Claude, and Gemini are built. Next-token prediction at scale, RLHF alignment, DPO, instruction tuning, and the laws that predict capability.
LLM Fine-Tuning in Practice
When to fine-tune vs RAG vs prompt. Full LoRA fine-tuning walkthrough on a real dataset using HuggingFace Transformers and PEFT.
Multimodal Models — CLIP, LLaVA, and Vision-Language
Models that see and understand images and text together. CLIP for zero-shot image classification, LLaVA for visual question answering.
Advanced RAG — Reranking, Hybrid Search and Evaluation
Beyond naive RAG. Reranking retrieved chunks, hybrid dense-sparse search, RAG evaluation metrics, and the patterns that separate production RAG from toy RAG.
Agents and Tool Use — Building Autonomous AI Systems
LLMs that plan, use tools, and execute multi-step tasks autonomously. ReAct, tool calling, memory, and the architecture patterns behind production AI agents.
ML Pipelines and Feature Stores
Feature pipelines, training pipelines, inference pipelines. Feast and Tecton for feature stores. Airflow, Kubeflow, and Prefect for orchestration.
Experiment Tracking with MLflow and Weights & Biases
Log every run, compare experiments, version models, register artifacts. Never lose a good experiment again.
Model Deployment — FastAPI, Docker, Kubernetes
Wrap your model in a FastAPI endpoint, containerise with Docker, scale with Kubernetes. Full working deployment of the Swiggy delivery model.
Model Monitoring — Drift Detection and Retraining
How to know your model is degrading before users complain. Data drift, concept drift, Evidently AI, and automated retraining triggers.
Retraining Pipelines — Keeping Models Fresh
Automated retraining schedules, trigger-based retraining on drift alerts, champion-challenger evaluation, and safe model promotion to production.
DVC — Data Version Control
Version datasets like code. DVC pipelines, remote storage, experiment tracking, and the full DVC + Git workflow for ML projects.
ML System Design — End to End
Design any ML system from scratch. The framework, trade-offs, capacity estimation, and how to present it in an interview.
Azure ML — Studio, Pipelines and AutoML
Azure Machine Learning Studio, compute clusters, AML Pipelines, AutoML, model registry, and online endpoints. Connects to the existing Azure track.
AWS SageMaker — Training Jobs and Pipelines
SageMaker training jobs, SageMaker Pipelines, Feature Store, Clarify for bias detection, and JumpStart model hub.
GCP Vertex AI — Pipelines and AutoML
Vertex Pipelines (Kubeflow-based), Feature Store, AutoML Tables, model endpoints, and BigQuery ML for SQL-native ML.
MLOps on Cloud — CI/CD for ML
GitHub Actions for ML pipelines, automated training and deployment, model validation gates, and the full MLOps loop on a cloud provider.
Interview Prep — 50 Complete ML Answers
50 complete answers covering classical ML, deep learning, system design, coding rounds, and behavioural questions — written at senior engineer depth.
Modules are dropping weekly.
Start with Module 01 the moment it goes live. Each module is self-contained enough to read on its own — but follow the order. Every concept earns the next one.
Discussion
0Have a better approach? Found something outdated? Share it — your knowledge helps everyone learning here.