Scan a chapter QR from the book (?from=book) or open a lab link — reader access unlocks after email. Direct visitors get a demo path with the same 20 labs.
System thinking for machine learning
Draw two diagrams for any ML system you have used or built: the model view and the system view. Compare both and identify where you are still operating as an operator rather than an architect.
Moving from leaderboard thinking to product outcomes
Take a competition-style ML task and rewrite it as a business system. Define the user, decision, metric, failure mode, and production feedback loop.
Turning raw data into useful model signal
Choose a dataset and design three feature groups: raw, transformed, and behavioural/contextual. Explain which feature could leak future information and how you would prevent it.
Understanding datasets through structure and distance
Run a PCA or UMAP-style analysis on a dataset and explain what the projection reveals, what it hides, and which clusters require additional investigation.
Leakage, bias, and misleading patterns
Audit a model or dataset for leakage and misleading correlation. Find one feature or segment where the model could look good in training and fail in reality.
Explaining model behaviour without pretending certainty
Prepare a model explanation for a technical reviewer and a non-technical stakeholder. Show what the explanation proves and what it does not prove.
Evaluation beyond one headline metric
Build an evaluation scaffold with global metrics, segment metrics, and failure examples. Find where the model performs acceptably overall but poorly for a specific slice.
Production readiness and operational thinking
Design a minimal production-ready ML pipeline: feature contracts, model registry, readiness gate, monitoring, and rollback path.
Choosing tools by system responsibility
Assemble a minimal ML stack for one project. Explain what handles tracking, serving, monitoring, data quality, and experiment review.
Retrieval as an evaluated system
Design a minimal RAG pipeline with chunking, embeddings, retrieval, and reranking. Define how you would measure Recall@5 and where the pipeline can fail.
Safe agent loops and control boundaries
Design a safe agent loop with hard limits, allow-list, logging, and fallback. Compare it with a simple prompt chain for the same task.
Cost, quality, and operational trade-offs
Compare a prompted baseline with a hypothetical fine-tuned variant. Define the evaluation protocol, expected quality gain, and ROI crossover point.
Latency, throughput, cost, and quality
Compare FP16, INT8, and INT4/AWQ as deployment choices. Explain the trade-off between speed, memory, and output quality for one use case.
Local, cloud, and hybrid compute choices
Estimate your GPU usage pattern and decide between buying, renting, or hybrid compute. Write a break-even logic for your case.
Provider risk, routing, and auditability
Design a data classification policy and gateway routing rules for an AI product. Include audit logging and a fallback path for provider changes.
Practical mathematical reasoning for ML engineers
Solve three practical reasoning tasks: tensor shapes in attention, A/B test power, and gradient debugging. Explain the engineering consequence of each result.
Data audit before model ambition
Run a data science audit: leakage, class imbalance, temporal split violations, missingness, and baseline sanity checks.
Algorithmic discipline inside ML systems
Translate an ML retrieval or ranking task into an algorithmic problem. Compare brute-force search with indexed search by latency and recall.
Structured ML system design communication
Take one ML system design prompt and answer it through six steps: objective, data, model, evaluation, serving, monitoring.
Turning the book into a concrete project path
Choose one project path after the book. Define the first two weeks: objective, dataset, prototype, evaluation, and evidence you will collect.
