Home / Artificial Intelligence / Artificial Intelligence
Artificial Intelligence
Advanced
v1.0.0
From the math underneath to production LLM systems.
A complete map of modern AI: the mathematics underneath every model, machine learning and deep learning fundamentals, NLP and computer vision, reinforcement learning, generative AI and large language models, and the MLOps, ethics, and career paths that surround production systems.
Your progress
0 / 55 topics
0 %
Saved in this browser. Clearing site data will reset it.
Topics in this roadmap
The interactive map needs JavaScript. Here is the full list.
Eigenvalues and Eigenvectors
Special vectors that only scale under a linear transformation.
Why this matters
Essential for PCA, spectral clustering, and understanding neural network dynamics.
Core concepts
Eigenvalues, eigenvectors, characteristic polynomial, eigendecomposition.
Practical skills
Compute eigenvalues using NumPy, apply PCA.
Where this is used in industry
Dimensionality reduction, recommendation systems, graph analysis.
Common mistakes
Assuming all matrices are diagonalizable.
Mini project
Implement PCA from scratch using eigendecomposition.
Interview importance: Medium — tested in ML theory questions.
Matrix Decompositions
Factoring matrices into canonical forms like SVD, LU, and QR decomposition.
Why this matters
SVD powers latent semantic analysis and collaborative filtering.
Core concepts
Singular Value Decomposition, QR decomposition, LU decomposition.
Practical skills
Apply SVD for image compression, use QR for least squares.
Where this is used in industry
Recommender systems, signal processing, solving linear systems.
Common mistakes
Using full SVD when truncated SVD is sufficient.
Mini project
Build an image compressor using truncated SVD.
Interview importance: Medium.
Derivatives and the Chain Rule
Rate of change of functions; gradients extend derivatives to multiple dimensions. The chain rule enables computing gradients through composed functions, which is the backpropagation algorithm.
Why this matters
Backpropagation is the single most important algorithm in deep learning.
Core concepts
Computational graph, forward pass, backward pass, local gradients.
Practical skills
Implement backpropagation from scratch for a simple neural network.
Where this is used in industry
All deep learning frameworks implement this.
Common mistakes
Gradient vanishing/exploding due to poor understanding.
Mini project
Build a mini autograd engine (like micrograd).
Interview importance: Very High — almost always asked in deep learning roles.
Probability and Bayes' Theorem
Mathematical formula for updating probabilities given new evidence.
Why this matters
Foundation of Naive Bayes classifiers, Bayesian inference, and reasoning under uncertainty.
Core concepts
Prior, likelihood, posterior, evidence, conjugate priors.
Practical skills
Build a Naive Bayes spam classifier.
Where this is used in industry
Spam filtering, medical diagnosis, A/B testing.
Common mistakes
Ignoring base rates (base rate fallacy).
Mini project
Build an email spam detector using Naive Bayes.
Interview importance: High.
Gradient-Based Optimization
Special class of optimization problems where local minima are global minima.
Why this matters
Theoretical guarantees for models like SVMs, linear regression.
Core concepts
Batch GD, SGD, Mini-batch GD, Momentum, RMSprop, Adam, AdamW, learning rate schedules.
Practical skills
Formulate problems as convex optimization.
Where this is used in industry
Portfolio optimization, SVM training, resource allocation.
Common mistakes
Assuming non-convex problems are convex.
Mini project
Solve a constrained optimization using CVXPY.
Interview importance: Low (general roles), High (research/algorithm roles).
Pandas for Data Manipulation
Data manipulation and analysis library.
Why this matters
Standard for data cleaning and exploration before modeling.
Core concepts
DataFrame, Series, groupby, merge, pivot, apply, handling missing data.
Practical skills
Clean messy datasets, perform exploratory data analysis.
Where this is used in industry
Any data science project starts here.
Common mistakes
Chained assignments, not using vectorized operations.
Mini project
Perform EDA on the Titanic dataset.
Interview importance: Very High for Data Science.
Pandas Documentation
Matplotlib and Seaborn Visualization
Data visualization libraries.
Why this matters
Visual communication of insights is as important as modeling.
Core concepts
Figure, Axes, plots (line, bar, scatter, histogram), customizations, subplots.
Practical skills
Create publication-ready figures.
Where this is used in industry
Exploratory analysis, model diagnostics, stakeholder presentations.
Common mistakes
Not labeling axes, using misleading scales.
Mini project
Visualize Titanic survival patterns.
Interview importance: Medium.
Object-Oriented Python for ML Code
Programming paradigm using classes and objects.
Why this matters
Essential for structuring large projects and many frameworks (PyTorch, Keras).
Core concepts
Classes, objects, inheritance, polymorphism, init, self, @classmethod, @staticmethod.
Practical skills
Design class hierarchies, inherit from framework base classes.
Where this is used in industry
Custom model definitions, PyTorch nn.Module.
Common mistakes
Overusing inheritance, not understanding self.
Mini project
Build a simple ML model base class with fit/predict methods.
Interview importance: Medium.
Version Control with Git
Track changes and collaborate on code.
Why this matters
Non-negotiable for any professional software role.
Core concepts
init, add, commit, push, pull, branch, merge, rebase, .gitignore.
Practical skills
Work in a team using feature branches, resolve merge conflicts.
Where this is used in industry
Every codebase.
Common mistakes
Committing large files, committing secrets.
Mini project
Contribute to an open-source AI project on GitHub.
Interview importance: Baseline expected.
Containerization with Docker
Package applications and dependencies into containers.
Why this matters
Solves "it works on my machine" and is standard for deployment.
Core concepts
Dockerfile, images, containers, docker-compose, volumes, registries.
Practical skills
Dockerize a Flask API serving a model.
Where this is used in industry
Model deployment, reproducible research environments.
Common mistakes
Large image sizes, running as root.
Mini project
Dockerize a trained model with a REST API.
Interview importance: High (MLOps/ML Engineer).
API Development for Model Serving
Building web interfaces for models.
Why this matters
Most models need to serve predictions over network.
Core concepts
REST, FastAPI, request/response, serialization, async.
Practical skills
Build a scalable model serving API.
Where this is used in industry
Model serving in production.
Common mistakes
No input validation, synchronous blocking.
Mini project
Deploy a sentiment analysis model as an API.
Interview importance: High.
Classification (Logistic Regression to SVMs)
Predicting discrete class labels for data points.
Why this matters
Applies to spam detection, image recognition, medical diagnosis, and customer churn prediction.
Core concepts
Decision boundary, log loss, hinge loss, sigmoid/softmax activation algorithms: Logistic Regression, SVM, Decision Trees, Random Forest, k-NN, Naive Bayes Practical skills
Build and tune classifiers using Scikit-learn; handle class imbalance.
Where this is used in industry
The most common supervised learning task in industry.
Common mistakes
Evaluating only on accuracy for imbalanced datasets.
Mini project
Build a multi-class classifier on the Iris or MNIST (non-DL) dataset.
Interview importance: Extremely High.
Clustering (K-Means, Hierarchical, DBSCAN)
Grouping data points into clusters based on similarity metrics without prior labels.
Why this matters
Critical for customer segmentation, anomaly detection, and data exploration when labels are absent.
Core concepts
Euclidean distance, centroids, density-based connectivity, evaluation metrics (silhouette score, inertia).
Practical skills
Apply K-Means, Agglomerative, and DBSCAN in Scikit-learn; select optimal number of clusters.
Where this is used in industry
Customer segmentation, image compression, anomaly detection.
Common mistakes
Not scaling features before clustering using K-Means on non-spherical data Mini project
Perform customer segmentation using real e-commerce RFM data.
Interview importance: High.
Dimensionality Reduction (PCA, t-SNE)
Compressing high-dimensional data into a lower-dimensional form while preserving essential structure.
Why this matters
Enables visualization, speeds up training, and mitigates the curse of dimensionality.
Core concepts
Variance retention, manifold hypothesis, reconstruction error.
Practical skills
Apply PCA and t-SNE; use autoencoders for non-linear reduction; visualize high-dimensional datasets.
Where this is used in industry
Feature preprocessing, noise reduction, 2D/3D data visualization.
Common mistakes
Using t-SNE for general preprocessing (only for visualization) not checking explained variance ratio Mini project
Visualize MNIST digits in 2D using PCA and t-SNE and compare them.
Interview importance: Medium-High.
Ensemble Methods (Bagging, Boosting, Stacking)
Bagging trains models in parallel on bootstrapped data; Boosting trains sequentially correcting errors; Stacking trains a meta-model on base model outputs.
Why this matters
These techniques dominate tabular data competitions and industry applications due to their high accuracy and robustness.
Core concepts
Bootstrapping, aggregation, weak learners, sequential error correction, residual fitting, level-0 vs level-1 models.
Practical skills
Implement Random Forest; tune XGBoost/LightGBM/CatBoost hyperparameters; build a stacking classifier.
Where this is used in industry
The default choice for structured/tabular data (fraud, churn, forecasting).
Common mistakes
Boosting without early stopping evaluating stacking without proper cross-validation leading to data leakage Mini project
Win a Kaggle tabular playground competition using an ensemble.
Interview importance: Extremely High.
XGBoost Documentation
Model Evaluation and Hyperparameter Tuning
Cross-validation simulates generalization; Bias-Variance Tradeoff explains error composition; Tuning navigates parameter space; Metrics define success criteria.
Why this matters
Transforms a notebook experiment into a trustworthy, production-ready evaluation. Errors here directly cause business failures.
Core concepts
k-Fold, Stratified, Time Series splits MSE vs RMSE vs MAE Confusion Matrix, F1, ROC-AUC, Precision-Recall curves Grid, Random, and Bayesian Search Practical skills
Set up Optuna pipelines; design time-series cross-validation; justify metric choice for imbalanced business cases.
Where this is used in industry
Model selection reports, backtesting trading algorithms, clinical trial validation.
Common mistakes
Normalizing before splitting (data leakage) using Random Search as a baseline without Bayesian optimization optimizing for accuracy with 99% negative class data Mini project
Build an ML pipeline with Scikit-learn Pipelines and Hyperopt/Optuna to beat a baseline.
Interview importance: Extremely High.
Feature Engineering
Selection removes noise; Extraction creates new latent variables; Encoding translates non-numeric entities into mathematics; Imputation reconstructs broken signals.
Why this matters
“Garbage in, garbage out.” Better features always beat more complex algorithms in real-world tabular data.
Core concepts
Chi-squared, Mutual Information, LASSO selection, Target Encoding, One-Hot Encoding, Mean/Median Imputation, MICE.
Practical skills
Implement Target Encoding with smoothing to prevent leakage; use Scikit-learn Transformers; build a Feature Store concept.
Where this is used in industry
80% of a Data Scientist’s time; fraud detection, credit scoring, recommendation engines.
Common mistakes
Applying Target Encoding on the full dataset Mean imputation without flagging the missingness pattern Mini project
Go from 80% to 90% accuracy on a Kaggle dataset purely by adding/cleaning features.
Interview importance: Critical (scenario-based interviews).
Optimization and Regularization
Optimizers adjust weights (SGD/Adam); Batch Norm stabilizes activations; Dropout prevents neuron co-adaptation; Early Stopping halts training at peak validation; Xavier/He Init sets initial weights.
Why this matters
Raw deep nets simply do not train well without these tricks. They determine training speed and generalization gap.
Core concepts
Adaptive learning rates (Adam/RMSProp), internal covariate shift, train vs test mode, gradient noise injection.
Practical skills
Fine-tune an AdamW schedule with warm restarts (Cosine Annealing); visualize learning curves to detect overfitting.
Where this is used in industry
Standard component of every training script.
Common mistakes
Applying Dropout during inference relying solely on Adam without examining learning rate plateaus Mini project
Train a model with and without Batch Norm; visualize the loss landscape difference.
Interview importance: High.
Convolutional Neural Networks
Convolutions extract spatial hierarchies; Pooling reduces resolution; Architectures (ResNet/EfficientNet) stack blocks; Transfer Learning adapts features; Detection localizes boxes; Segmentation classifies pixels.
Why this matters
Computer Vision is the first grand success of Deep Learning. These techniques power autonomous driving, medical imaging, and AR.
Core concepts
Filters/kernels, stride, padding, receptive field, residual connections, IoU, mAP, FPN, U-Net.
Practical skills
Fine-tune a ResNet-50 on a custom dataset in PyTorch; implement a YOLO inference pipeline; build a U-Net for medical imaging.
Where this is used in industry
Quality control, satellite imagery analysis, facial recognition.
Common mistakes
Applying pre-trained Imagenet weights to drastically different modalities without adaptation incorrect mAP calculation Mini project
Train a trash classifier on the TACO dataset; detect pedestrians on video.
Interview importance: High.
CS231n: CNNs for Visual Recognition
Recurrent Networks (RNN, LSTM, GRU)
RNNs loop over sequences; LSTMs introduce forget/input/output gates to combat vanishing gradients; Seq2Seq uses encoder-decoder for translation.
Why this matters
While Transformers have replaced RNNs in many areas, RNNs remain fundamental for understanding attention and efficient on-device streaming models.
Core concepts
Hidden state recurrence, vanishing gradients, forget gate, cell state, attention (Bahdanau/Luong).
Practical skills
Build a character-level text generator; train an attention-based Seq2Seq model for date translation.
Where this is used in industry
Time-series forecasting, audio generation, legacy NLP systems.
Common mistakes
Forgetting to reset hidden state between epochs applying RNNs to very long sequences without truncation/attention Mini project
Build a Shakespeare text generator using an LSTM.
Interview importance: Medium (Legacy knowledge; Transformers are taking over).
Transformers and Attention
Attention computes weighted representations; Transformer stacks Self-Attention and FFN layers; BERT uses Encoder for understanding; GPT uses Decoder for generation; ViT splits images into patches.
Why this matters
Transformers have become the dominant architecture in NLP, Vision, and Multi-Modal AI. "Attention is All You Need."
Core concepts
Query/Key/Value, Multi-Head Attention, Positional Encoding, Encoder/Decoder masks, Cross-Attention, Pre-training vs Fine-tuning.
Practical skills
Fine-tune a BERT model on a GLUE task (Hugging Face); generate text using a GPT-2 checkpoint; train a ViT on CIFAR-10.
Where this is used in industry
Search engines (Google), chatbots, code generation (Copilot), protein folding.
Common mistakes
Not handling padding correctly in attention masks using standard attention for very long sequences without optimization Mini project
Build a question-answering bot using fine-tuned BERT on SQuAD.
Interview importance: Extremely High.
The Illustrated Transformer
Attention Is All You Need (paper)
Generative Models (GANs, VAEs, Diffusion)
GANs pit generator vs discriminator; VAEs maximize ELBO; Diffusion Models reverse a gradual noising process; Flow Models apply invertible transformations.
Why this matters
The engine behind DALL-E, Stable Diffusion, and deepfakes. Critical for creative AI, data augmentation, and simulation.
Core concepts
Min-max game, KL divergence, Reparameterization trick, Markov chains, noise scheduling, latent space interpolation.
Practical skills
Generate synthetic faces (StyleGAN); train a Conditional Diffusion Model on custom data; interpolate in latent space.
Where this is used in industry
Creative design, drug discovery, synthetic data generation.
Common mistakes
Unstable GAN training (non-saturating loss ignored) judging sample quality by eye only without FID/Inception Score Mini project
Build a DCGAN to generate MNIST digits; implement a Denoising Diffusion Probabilistic Model (DDPM) simplified.
Interview importance: High.
Text Preprocessing (Tokenization, Stemming, NER)
Tokenization splits text; Stemming/Lemmatization reduce words to roots; Stop Words remove noise; POS identifies grammar roles; NER extracts entities.
Why this matters
The quality of NLP model input directly limits model performance. Garbage text = garbage embeddings.
Core concepts
Regex tokenizers, BPE/WordPiece, Snowball/Porter stemmer, WordNet lemmatizer, Penn Treebank tags, IOB format.
Practical skills
Build a preprocessing pipeline with SpaCy; perform custom entity extraction with Regex.
Where this is used in industry
Search engines, document processing, legal tech.
Common mistakes
Removing stop words for sentiment analysis (alters meaning) heavy stemming breaking words Mini project
Build a CV parser extracting names, emails, and skills using NER.
Interview importance: Medium.
Word Embeddings (Word2Vec, GloVe, Contextual)
Word2Vec predicts context; GloVe factorizes co-occurrence; FastText uses subword info; Contextual embeddings (ELMo/BERT) change based on context.
Why this matters
Understanding embeddings is key to understanding how models "see" language.
Core concepts
CBOW, Skip-gram, negative sampling, cosine similarity, word analogies.
Practical skills
Train Word2Vec on a custom corpus (Gensim); use pre-trained GloVe; query nearest neighbors.
Where this is used in industry
Information retrieval, recommendation, feature representation.
Common mistakes
Using static embeddings for polysemous words (e.g., "bank") without context.
Mini project
Train a domain-specific embedding and evaluate with analogy tasks.
Interview importance: Medium.
Core NLP Tasks (Classification, Translation, QA)
Classification assigns labels; Sentiment identifies polarity; Translation converts languages; QA extracts answers; Summarization condenses text (Extractive/Abstractive).
Why this matters
These are the direct business applications of NLP in production.
Core concepts
BLEU/ROUGE/METEOR metrics, Beam Search, Encoder-Decoder models, Span prediction.
Practical skills
Fine-tune T5 for summarization; build a multilingual sentiment analyzer; deploy a QA bot using Haystack.
Where this is used in industry
Customer feedback, product reviews, document automation.
Common mistakes
Using greedy decoding instead of beam search for translation evaluating summarization only on ROUGE Mini project
Build a multi-lingual translation API using MarianMT.
Interview importance: Very High.
Deep Learning for Vision (Detection, Segmentation)
Classification labels an image; Detection localizes multiple objects with boxes; Segmentation classifies every pixel.
Why this matters
Core applied CV tasks. Detection and Segmentation are the primary deliverables in surveillance, autonomous driving, and medical diagnosis.
Core concepts
Anchor boxes, Non-Max Suppression (NMS), Feature Pyramid Networks (FPN), DICE loss, Intersection over Union (IoU).
Practical skills
Train a YOLOv8 model on a custom dataset (Ultralytics); deploy a Mask R-CNN model for defect detection.
Where this is used in industry
Retail (shelf analytics), Manufacturing (crack detection), Biotech (cell counting).
Common mistakes
Using a pre-trained model without fine-tuning on target domain aspect ratios incorrect bounding box format (xyxy vs xywh) Mini project
Train a custom pet/dog breed detector and deploy it on a mobile video feed.
Interview importance: Very High.
Video Analysis (Tracking, Action Recognition)
Motion detection identifies changes; Tracking maintains identity (ID) across frames; Action Recognition classifies temporal sequences (3D CNNs/Transformers).
Why this matters
Essential for surveillance, sports analytics, and self-driving cars.
Core concepts
Optical Flow, Background Subtraction, Intersection over Union (IoU) matching, Kalman Filters, Skeleton-based action recognition.
Practical skills
Build a people counter for a mall; implement a DeepSORT tracker on top of YOLO; train a TimeSformer model.
Where this is used in industry
Traffic management, athletics performance analysis, anomaly detection.
Common mistakes
Naive frame-by-frame detection without MOT (Multiple Object Tracking) integration.
Mini project
Track a soccer ball and players in a match clip.
Interview importance: Medium.
3D Vision (Point Clouds, NeRF)
Reconstruction creates 3D models; Point Clouds are sparse 3D representations; Depth maps measure pixel distance; NeRF renders photorealistic scenes.
Why this matters
The next frontier for AR/VR, robotics simulation, and digital twins (Metaverse).
Core concepts
Epipolar geometry, PnP, SfM, voxels, LiDAR processing, volumetric rendering.
Practical skills
Generate a depth map from a stereo pair; train a tiny NeRF model on a custom scene; process a point cloud with Open3D.
Where this is used in industry
Warehouse automation, digital heritage preservation, VR content creation.
Common mistakes
Underestimating calibration requirements.
Mini project
Capture a toy figure with your phone and reconstruct it using 3D Gaussian Splatting.
Interview importance: Medium (Growing rapidly).
Deep RL (DQN, Policy Gradients, PPO)
DQN approximates Q-values; Policy Gradients optimize policy directly; Actor-Critic combines both (A2C/A3C); PPO constrains policy updates for stability.
Why this matters
This is the practical standard for applying RL to real problems. PPO is widely used in OpenAI and DeepMind for robotics.
Core concepts
Experience Replay, Target Networks, Advantage Function, Surrogate Objective, Clipping.
Practical skills
Solve CartPole and LunarLander using Stable-Baselines3; write a PPO training loop from scratch.
Where this is used in industry
Autonomous drone racing, dexterous hand manipulation, NPC behavior in games.
Common mistakes
Not standardizing state vectors PPO without clipping leads to catastrophic policy updates Mini project
Train an agent to play the Flappy Bird clone using DQN.
Interview importance: Very High.
Advanced RL (Multi-Agent, Model-Based)
Multi-Agent RL handles cooperation/competition; Inverse RL infers reward from demonstrations; Model-Based RL learns the transition dynamics; Exploration addresses the explore-exploit dilemma (Intrinsic motivation).
Why this matters
Represents the cutting-edge of RL research, aiming for sample efficiency and human-like learning.
Core concepts
Nash Equilibrium in RL, Adversarial Imitation, World Models, Curiosity, RND.
Practical skills
Train a self-play agent in a simple game; use GAIL to imitate an expert policy; implement an MPC controller with a learned model.
Where this is used in industry
Game balancing, autonomous driving planning, energy grid optimization.
Common mistakes
Forgetting to normalize intrinsic rewards ignoring catastrophic forgetting in model-based learning Mini project
Implement a Curiosity-driven agent to explore a maze without external rewards.
Interview importance: Medium (Research focused).
Retrieval-Augmented Generation
Combines a language model with a retrieval step over an external knowledge base, so the model answers using documents fetched at query time instead of relying only on what it memorised during training.
Why this matters
RAG is the standard way to ground an LLM in current, private, or domain-specific data without the cost of fine-tuning, and it reduces hallucination by giving the model something concrete to cite.
Core concepts
Embedding documents into a vector store Similarity search / nearest neighbour retrieval Chunking strategy and its effect on answer quality Re-ranking retrieved passages before feeding them to the model Common mistakes
Chunking documents too large or too small, hurting retrieval precision Not evaluating retrieval quality separately from generation quality Interview importance: High, RAG system design is one of the most common practical LLM interview questions.
LangChain RAG Docs
Fine-tuning and Instruction Tuning
Adapts a pretrained model's weights on a smaller, task-specific dataset. Instruction tuning specifically trains the model to follow natural language instructions rather than just complete text.
Why this matters
Fine-tuning gets a general model to behave reliably on a narrow task, or to adopt a specific tone or format, more consistently than prompting alone can achieve.
Core concepts
Full fine-tuning vs parameter-efficient methods (LoRA, QLoRA) Instruction tuning datasets and formats Catastrophic forgetting and how to avoid it Interview importance: High for LLM application and ML engineering roles.
Hugging Face PEFT (LoRA)
Model Quantization and Pruning
Reduces a trained model's size and inference cost, usually with a small accuracy tradeoff, so it can run faster or on smaller hardware.
Core concepts
Post-training quantization vs quantization-aware training Structured vs unstructured pruning Knowledge distillation into a smaller student model Why this matters
Inference cost, not training cost, dominates the lifetime expense of most deployed models, especially at high traffic or on edge devices.
Interview importance: Medium to high for MLOps and platform engineering roles.
Data Pipelines for AI
The infrastructure that gets raw data into a state a model can train on: batch and streaming pipelines, feature stores, and data versioning so experiments stay reproducible.
Core concepts
Batch vs streaming ingestion Feature stores for sharing features between training and serving Data versioning and lineage Interview importance: Medium, more central to MLOps and data engineering roles than to research roles.
Adversarial Attacks and Model Security
Small, often imperceptible perturbations to input data can cause a model to misclassify with high confidence. Model security also covers extraction and membership inference attacks that leak information about the training data.
Core concepts
Adversarial examples and perturbation budgets Adversarial training as a defence Model extraction and membership inference attacks Interview importance: Growing, especially for AI safety and security focused roles.
AI Governance Frameworks
The organisational and regulatory processes around deploying AI responsibly: model documentation, audit trails, and compliance with emerging regulation such as the EU AI Act.
Core concepts
Model cards and datasheets for datasets Audit trails and explainability requirements Regulatory landscape: EU AI Act, sector-specific rules Interview importance: Growing, particularly at larger companies and in regulated industries.
Graph Neural Networks
Extends deep learning to graph-structured data, learning representations that respect a graph's connectivity rather than treating inputs as an unordered vector or grid.
Core concepts
Message passing between connected nodes Graph convolutional and graph attention layers Applications in molecule and drug discovery Interview importance: Niche but growing, particularly in biotech and recommendation-heavy companies.
Time Series Forecasting
Models that predict future values from historical sequences, from classical statistical methods through to deep learning approaches built for temporal data.
Core concepts
ARIMA and classical statistical forecasting Prophet for business time series with seasonality Deep learning approaches (temporal convolutions, transformers for time series) Interview importance: Medium, common in finance, demand forecasting, and operations roles.
Recommendation Systems
Systems that predict which items a user is most likely to want next, combining signals about the user, the item, and their past interactions.
Core concepts
Collaborative filtering vs content-based filtering Hybrid systems combining both signal types Two-tower architectures for retrieval at scale Interview importance: High at any company with a consumer product, one of the most commonly asked system design topics.
Machine Learning Engineer — $120K to $200K
Bridges software engineering and data science, building and deploying scalable models. Requires Python, TensorFlow/PyTorch, SQL, MLOps, Docker, and Kubernetes. Portfolio needs 2-3 end-to-end deployed projects with a strong GitHub and optimised inference pipelines.
Data Scientist — $110K to $190K
Analyses data, builds models, and generates insights that drive business decisions. Requires Python (Pandas/Scikit-learn), SQL, statistics, visualisation, and experimentation (A/B testing). Interview prep centres on product sense, statistical modelling, and SQL.
AI Research Scientist — $150K to $250K+
Invents novel architectures and algorithms, publishes papers, and pushes the state of the art. Often requires a PhD, deep JAX/PyTorch internals knowledge, and heavy math. Portfolio value comes from top-tier publications (NeurIPS, ICML, ICLR, CVPR).
NLP / Computer Vision Engineer — $130K to $210K
Specialist applying domain-specific architectures to unstructured text or images. Requires Transformers, Hugging Face, OpenCV or Diffusers, and domain data augmentation skill.
MLOps / Platform Engineer — $140K to $220K
Builds the infrastructure and automation that lets other ML teams ship. Requires Terraform, AWS or GCP, Kubernetes, CI/CD, Python, and monitoring tooling.
Vectors and Matrices
Fundamental building blocks representing data points and linear transformations.
Why this matters
All data in AI is represented as vectors and matrices.
Core concepts
Vector addition, scalar multiplication, dot product, matrix multiplication, transpose, inverse.
Practical skills
Implement vectorized operations in NumPy.
Where this is used in industry
Data representation, feature engineering, neural network layers.
Common mistakes
Confusing dot product with element-wise multiplication.
Mini project
Build a simple image transformation tool using matrix operations.
Interview importance: High — foundational for all ML interviews.
NumPy Linear Algebra
3Blue1Brown: Essence of Linear Algebra
MIT 18.06 Linear Algebra (OCW)
NumPy
Library for large, multi-dimensional arrays and matrices.
Why this matters
Backbone of nearly all scientific computing in Python.
Core concepts
ndarray, broadcasting, vectorization, slicing, ufuncs.
Practical skills
Replace slow loops with vectorized operations, perform linear algebra.
Where this is used in industry
Input preprocessing, image manipulation, model implementation from scratch.
Common mistakes
Misunderstanding broadcasting rules.
Mini project
Build a simple image filter using NumPy.
Interview importance: High.
NumPy Documentation
Regression
Predicting continuous numerical values.
Why this matters
The most fundamental predictive modeling task; underpins forecasting, pricing, and risk models.
Core concepts
Least squares, loss functions (MSE, MAE), polynomial features, regularization (L1/L2).
Practical skills
Implement and train regressors using Scikit-learn; interpret coefficients.
Where this is used in industry
Housing prices, stock forecasting, sales predictions.
Common mistakes
Extrapolating beyond data range, ignoring multicollinearity.
Mini project
Predict house prices using the Boston Housing or California Housing dataset.
Interview importance: Very High.
Scikit-learn: Linear Models
Neural Network Fundamentals
A perceptron is a single neuron model. Activation functions induce non-linearity. Forward Prop computes predictions; Backprop computes gradients; Loss measures penalty.
Why this matters
This is the atomic unit of Deep Learning. Understanding these components is the difference between a framework user and an engineer who can debug convergence failures.
Core concepts
Weighted sums, biases, sigmoid/ReLU/tanh dynamics, vanishing gradients, Cross-Entropy vs MSE loss landscapes.
Practical skills
Build a mini framework similar to PyTorch autograd (micrograd); visualize decision boundaries for different activations; numerically verify gradient correctness.
Where this is used in industry
Fundamental block of every deep learning system.
Common mistakes
Using Sigmoid in hidden layers (saturation) forgetting to zero gradients before backward step Mini project
Build a two-layer neural network from scratch to classify MNIST digits (>95% accuracy).
Interview importance: Extremely High.
Deep Learning Book (Goodfellow et al.)
LLM Applications and Deployment
Prompt Engineering structures inputs; Fine-tuning adapts weights (LoRA); RAG grounds generation in external documents; Deployment optimizes serving (vLLM); Evaluation benchmarks bias/truthfulness.
Why this matters
The defining skill of the current generation of AI engineers. Underpins ChatGPT and enterprise LLM products.
Core concepts
Few-shot/Chain-of-Thought prompting, PEFT, Adapters, Retrievers (BM25/Dense), Hallucination, Embedding models.
Practical skills
Build a "Chat with your PDF" app using LangChain/LlamaIndex; fine-tune Llama-3 with LoRA on a custom dataset; quantize and serve a model with Ollama.
Where this is used in industry
Customer support automation, knowledge management, code copilots.
Common mistakes
Not chunking documents appropriately for RAG over-relying on GPT-4 without evaluating smaller fine-tuned models Mini project
Create a domain-specific Q&A bot using LangChain, ChromaDB, and a local open-source LLM.
Interview importance: Extremely High.
Hugging Face NLP Course
Prompt Engineering Guide
Image Processing Basics
Filtering smooths/sharpens; Edge Detection finds boundaries (Canny); Segmentation partitions regions (Watershed); Feature Detection finds keypoints (SIFT/ORB).
Why this matters
Deep learning isn't always the answer. These classical algorithms provide lightweight, interpretable tools for preprocessing and industrial inspection.
Core concepts
Kernels, Gaussian blur, Canny thresholds, connected components, corner detection.
Practical skills
Implement a document scanner app using morphological ops; stitch panoramas using feature matching (OpenCV).
Where this is used in industry
Barcode/QR scanning, robotic pick-and-place, document OCR.
Common mistakes
Using SIFT in performance-critical real-time apps without GPU alternatives.
Mini project
Build a real-time lane detection system for a dashcam video using edge detection.
Interview importance: Medium.
OpenCV Documentation
Reinforcement Learning Foundations
MDPs define states/actions/rewards; Value Functions predict return; Policy Iteration alternates between evaluation and improvement; Q-Learning learns action-values from experience.
Why this matters
The Bellman Equation is the cornerstone of all modern RL.
Core concepts
Agent, Environment, Reward Hypothesis, Bellman Optimality Equation, Temporal Difference Learning.
Practical skills
Implement a Gridworld solver; apply Tabular Q-Learning to solve FrozenLake (OpenAI Gym).
Where this is used in industry
Operations research, queuing theory, basic game logic.
Common mistakes
Mismatching Q-Learning (off-policy) and SARSA (on-policy) update rules.
Mini project
Train an agent to navigate a maze using Q-Learning.
Interview importance: High.
Sutton and Barto: Reinforcement Learning
Generative AI Foundations
Prompt Engineering manipulates model conditioning; RLHF aligns models to human values; Stable Diffusion models images; ControlNet adds spatial guidance; Text-to-Speech generates audio.
Why this matters
Core skill for building next-gen products. Companies urgently need engineers who can control these models.
Core concepts
Classifier-free guidance, LoRA for diffusion, CLIP space, latent diffusion, adversarial prompt injections.
Practical skills
Design an InstructGPT-style fine-tuning data format; train a Dreambooth model on personal photos; set up an automatic voice-over pipeline.
Where this is used in industry
Ad creative generation, game asset creation, audiobook narration, deepfake detection.
Common mistakes
Deploying models without guardrails prompt injection vulnerabilities Mini project
Fine-tune a Stable Diffusion model to generate custom avatars; build a Telegram bot that generates images.
Interview importance: Extremely High.
OpenAI Cookbook
Stable Diffusion Docs
ML Infrastructure and Serving
Tracking logs metrics/params; Versioning stores artifacts; Serving handles inference; Kubernetes orchestrates scaling; Feature Stores standardize features.
Why this matters
Without MLOps, 87% of ML projects never reach production. This is the "last mile" of AI value delivery.
Core concepts
Pipeline reproducibility, CI/CD for ML, online vs batch inference, cold starts, GPU sharing, training-serving skew.
Practical skills
Set up an MLflow server; optimize a model with TensorRT/ONNX; deploy a model behind a gRPC endpoint scaled via K8s HPA.
Where this is used in industry
Fintech real-time fraud detection, e-commerce recommendation API.
Common mistakes
Monitoring only system metrics, not data drift high latency due to missing model optimizations Mini project
Build an end-to-end pipeline: GIT push -> Retrain model -> Deploy if accuracy improves.
Interview importance: Very High.
MLflow Documentation
Kubernetes Documentation
Fairness and Explainability
Fairness requires measuring disparate impact; Explainability uses game theory to assign feature importance; Privacy limits information leakage; Governance codifies risk management.
Why this matters
Failing these concepts results in lawsuits, fines, and catastrophic public trust failures.
Core concepts
Demographic parity, equalized odds, SHAP values, epsilon budget, data minimization, EU AI Act.
Practical skills
Run a fairness audit using the What-If Tool; explain a loan default model using SHAP; train a Differentially Private ML model (Opacus).
Where this is used in industry
Loan underwriting, resume screening, medical diagnostics.
Common mistakes
Ignoring intersectional bias treating explainability as a post-hoc checkbox Mini project
Take a biased dataset and show the performance metrics before/after applying a bias mitigation technique.
Interview importance: High (Leadership and Ethics rounds).
SHAP Documentation
Specialized AI Domains
VLMs align images and text (CLIP); GNNs learn on graphs; Time Series models predict futures; Two-Tower architectures serve large-scale Retrieval in RecSys.
Why this matters
These techniques power GPT-4V, AlphaFold, stock prediction, and TikTok's "For You" page.
Core concepts
Contrastive loss (InfoNCE), message passing, neighbor sampling, stationarity, autoregression, embedding lookups, cosine similarity search.
Practical skills
Build a semantic search engine using CLIP embeddings; predict energy consumption with Prophet; train a GNN for molecular property prediction.
Where this is used in industry
Drug discovery, financial engineering, content discovery.
Common mistakes
Applying GNNs without feature normalization using ARIMA for long-horizon volatile series Mini project
Implement a multimodal search (text->image) for your local photo album.
Interview importance: High for specialist roles.
