Machine
Learning.

F1 · Beginner · 75 min

Machine learning is applied math. You do not need a PhD, but you do need the core: linear algebra (vectors, matrices), calculus (gradients — how models learn), and probability (uncertainty and likelihood). These three show up in every model you will ever train.

• Vectors & matrices
• Gradients & derivatives
• Probability & distributions
• Dot products & norms

Build itCompute a gradient by hand for a tiny function, then explain what "descending the gradient" means for a model chasing lower loss.

✓ CheckpointExplain what a gradient tells you geometrically, and why that is enough to train anything at all.

F2 · Beginner · 60 min

Python is the language of ML, and NumPy/pandas are its workhorses. Learn vectorized array math (loops are the enemy of speed), dataframes for tabular data, and the scikit-learn API shape. Fluency here is what lets you go from idea to experiment in minutes.

• NumPy arrays & vectorization
• pandas dataframes
• scikit-learn API
• Plotting & EDA

Build itVectorize a slow Python loop with NumPy and measure the speedup. Why is the array version often 100x faster?

✓ CheckpointExplain why vectorised NumPy beats a Python loop by so much that the difference changes what is feasible.

F3 · Intermediate · 65 min

Models are only as good as their features. Learn cleaning, handling missing values, encoding categoricals, scaling, and crafting features that expose signal. In classical ML, thoughtful feature engineering beats a fancier model more often than not.

• Cleaning & missing values
• Encoding categoricals
• Scaling & normalization
• Feature creation & selection

Build itTurn a raw table with dates, categories, and nulls into model-ready features. Which engineered feature would most help a churn model?

✓ CheckpointExplain why fitting your scaler on the full dataset before splitting leaks the test set — and how the leak shows up as an optimistic score.

F4 · Intermediate · 70 min

The dominant paradigm: learn a mapping from inputs to labeled outputs. Learn the split between classification and regression, training vs test data, the bias-variance tradeoff, and overfitting. This mental model underpins nearly everything else in the roadmap.

• Classification vs regression
• Train / validation / test
• Bias-variance tradeoff
• Overfitting & regularization

Build itA model scores 99% on training and 70% on test. Name the problem, and three ways to close the gap.

✓ CheckpointState the bias-variance position of a model that scores 99% on train and 70% on test, and what you would change.

F5 · Intermediate · 55 min

When there are no labels, structure must be found in the data itself. Learn clustering (k-means), dimensionality reduction (PCA), and embeddings — the tools for exploration, compression, and finding hidden groups. This is also where representation learning begins.

• Clustering (k-means)
• Dimensionality reduction (PCA)
• Embeddings
• Anomaly detection

Build itYou have customers but no segments. Cluster them, then reduce to 2D to visualize. How do you decide how many clusters are real?

✓ CheckpointExplain how you would judge whether k-means found real structure or just partitioned noise.

F6 · Intermediate · 60 min

A single accuracy number lies. Learn the metrics that matter — precision, recall, F1, ROC-AUC for classification; MAE/RMSE for regression — cross-validation, and why the right metric depends on the cost of each kind of error. Evaluation is where honesty lives.

• Precision / recall / F1
• ROC-AUC & confusion matrix
• Cross-validation
• Choosing the right metric

Build itA fraud model is 99.9% accurate but useless. Explain why, and which metric you should have optimized instead.

✓ CheckpointExplain why accuracy is the wrong metric on a 1%-positive dataset, and which metric you would report instead.

T1 · Intermediate · 60 min

Start where ML started. Linear regression fits a line; logistic regression bends it into a probability for classification. Learn the loss functions, gradient descent, and regularization. These simple, interpretable models are still the right first answer to a huge share of problems.

• Linear regression
• Logistic regression
• Loss & gradient descent
• L1 / L2 regularization

Build itFit logistic regression to a binary problem, then read the coefficients. What does a large positive weight tell you about a feature?

✓ CheckpointExplain what logistic regression actually outputs, and why calling it a probability requires calibration.

T2 · Advanced · 65 min

Decision trees split data into rules; ensembles combine many weak trees into a strong model. Learn random forests (bagging) and gradient boosting (XGBoost/LightGBM). On tabular data, boosted trees remain the models to beat — often outperforming deep nets.

• Decision trees
• Random forests (bagging)
• Gradient boosting (XGBoost)
• Feature importance

Build itA boosted-tree model wins on your tabular data. Explain why trees often beat neural nets here — and what feature importance reveals.

✓ CheckpointExplain the difference between bagging and boosting in terms of what each one is trying to reduce.

T3 · Advanced · 75 min

Neural nets stack simple units into powerful function approximators. Learn neurons, activation functions, forward passes, and backpropagation — the algorithm that assigns credit and lets a network learn. Understand backprop once and every deep model stops being magic.

• Neurons & layers
• Activation functions
• Forward pass
• Backpropagation

Build itTrace one training step through a 2-layer net: forward to a loss, then backprop the gradient. Where does the chain rule appear?

✓ CheckpointExplain what backpropagation is doing with the chain rule, and why the forward pass has to be remembered.

T4 · Advanced · 75 min

Depth unlocks representation learning — features are learned, not hand-crafted. Learn training at scale: optimizers (Adam), batching, learning-rate schedules, regularization (dropout), and the frameworks (PyTorch) that make it practical. This is the gateway to modern AI.

• Optimizers (SGD/Adam)
• Batching & epochs
• Dropout & batch norm
• PyTorch basics

Build itYour deep net will not converge. Walk through learning rate, initialization, and normalization as the usual suspects — which do you check first?

✓ CheckpointExplain what dropout does at training time and what it must NOT do at inference.

T5 · Advanced · 65 min

Convolutional networks brought the deep-learning revolution to images. Learn convolutions, pooling, feature hierarchies, and landmark architectures (ResNet). Even in a transformer world, CNNs remain the efficient default for many vision tasks — and the intuitions transfer.

• Convolutions & filters
• Pooling & feature maps
• Architectures (ResNet)
• Transfer learning

Build itFine-tune a pretrained ResNet on a small image dataset. Why does transfer learning beat training from scratch with little data?

✓ CheckpointExplain what a convolution buys over a fully-connected layer for images — the answer is about structure, not size.

T6 · Advanced · 75 min

Transformers replaced recurrence with attention and now dominate language — and beyond. Learn tokenization, embeddings, self-attention, and the encoder/decoder split that underpins BERT and GPT. This is the architecture behind every modern large language model.

• Tokenization & embeddings
• Self-attention
• Encoder vs decoder
• Pretraining & fine-tuning

Build itExplain self-attention as "each token looks at every other token." Why does that beat a fixed-window model for long-range meaning?

✓ CheckpointExplain why self-attention is quadratic in sequence length, and what that costs on a long document.

P1 · Advanced · 50 min

ML without tracking is chaos — dozens of runs and no memory of what worked. Learn experiment tracking (metrics, params, artifacts), data and model versioning, and reproducibility. The goal: any result can be reproduced, and the best model is never a mystery.

• Experiment tracking
• Data & model versioning
• Reproducibility
• Config management

Build itYou beat your baseline last week but cannot reproduce it. What should you have logged — data version, seed, params — to get it back?

✓ CheckpointExplain why an experiment you cannot reproduce is not a result yet.

P2 · Advanced · 65 min

A model in a notebook helps no one. Learn to serve models — batch vs real-time inference, packaging, APIs, and CI/CD for models. MLOps brings software discipline to ML so models ship reliably and repeatedly, not once by hand.

• Batch vs real-time serving
• Model packaging & APIs
• CI/CD for models
• Rollout & rollback

Build itTurn a trained model into a low-latency API. Where do you validate inputs, and how do you roll back a bad model version fast?

✓ CheckpointExplain what has to be true of a model before it can be served, beyond the metric it was selected on.

P3 · Advanced · 50 min

Features used in training must match features at serving time, or the model silently breaks. A feature store centralizes feature definitions, serves them online and offline, and prevents training-serving skew — a subtle bug that quietly wrecks production models.

• Online vs offline features
• Training-serving skew
• Feature reuse
• Point-in-time correctness

Build itA model is great offline but poor live. Show how a feature computed differently in training vs serving causes it — and how a store fixes it.

✓ CheckpointExplain training-serving skew with a concrete example, and how a feature store prevents it.

P4 · Advanced · 60 min

Models decay because the world changes. Learn to monitor prediction quality, detect data drift and concept drift, and trigger retraining. Unlike a service that stays broken until fixed, a model degrades quietly — monitoring is what catches the slow rot.

• Data vs concept drift
• Prediction monitoring
• Retraining triggers
• Shadow & canary models

Build itAccuracy is fine but the input distribution shifted. Which drift metric fires first, and when does that justify a retrain?

✓ CheckpointExplain the difference between data drift and concept drift, and why the second is harder to detect.

P5 · Advanced · 60 min

Big models and big data outgrow one GPU. Learn data parallelism vs model parallelism, gradient synchronization, mixed precision, and the memory/throughput tradeoffs. This is how modern models are trained across many accelerators without falling over.

• Data vs model parallelism
• Gradient synchronization
• Mixed precision
• GPU memory & throughput

Build itA model no longer fits on one GPU. Choose data vs model parallelism and explain the communication cost each strategy adds.

✓ CheckpointExplain the difference between data and model parallelism in terms of what gets split across devices.

P6 · Advanced · 55 min

Powerful models carry real risk. Learn fairness and bias auditing, interpretability (SHAP, feature attribution), privacy, and the human oversight that keeps automated decisions accountable. Responsible AI is not a compliance afterthought — it is part of the engineering.

• Fairness & bias auditing
• Interpretability (SHAP)
• Privacy & data ethics
• Human oversight

Build itA loan model rejects one group more often. How do you measure whether it is biased, and what levers reduce the disparity?

✓ CheckpointExplain what SHAP tells you and what it does not — an explanation is not a justification.