ML in Practice
Feature engineering, deployment pitfalls, and algorithms from scratch — bridging the gap between theory and production ML.
22 lessons
01Feature Engineering
How to scale, encode, and combine raw features for model-ready data — plus the leakage traps that silently corrupt your evaluation.
Deployment Pitfalls
Train-serve skew, data drift, monitoring, calibration, and interpretability — everything that goes wrong after the notebook closes.
Algorithms from Scratch
Interview-style implementations that reveal the fundamentals hidden inside ML libraries — assumptions, edge cases, and the math beneath the API.
ML Systems Design & Problem Framing
Before you train a model, frame the problem: reliability, scalability, maintainability, the business objective behind your loss, and the gap between research and production.
Data Engineering Fundamentals
Where ML data actually comes from and how it is stored: row vs column formats, OLTP vs OLAP, ETL vs ELT, and batch vs streaming pipelines.
Training Data: Sampling, Labelling, Imbalance
How to construct a training set that does not lie to your model: sampling strategies, labelling at scale, and the canonical class-imbalance toolbox.
Experiment Tracking & Versioning
Runs, metrics, artifacts, and the three things you must version — code, data, environment — to make a training run reproducible months later.
Deployment Patterns & Model Compression
Batch, online, and edge serving — and the compression stack (quantization, pruning, distillation) that makes each pattern affordable in production.
Continual Learning & Testing in Production
How models stay fresh after launch — stateless vs stateful retraining, distribution shift triggers, and the four canonical traffic-split patterns (shadow, canary, A/B, bandits) for testing changes against live users.
Monitoring & Observability
Why ML monitoring needs four layers — system, input, output, outcome — and how percentile latencies, PSI drift, calibration drift, and SLO error-budget burn rates compose into alerts that wake the right person at the right time.
MLOps Infrastructure & Orchestration
Storage and compute layers, development environments, resource management, workflow orchestrators (Airflow/Prefect/Dagster, Argo, Kubeflow), and what an ML platform actually buys you.
Inference Optimization & Serving
Make trained models fast and cheap to serve — KV-cache for attention, continuous batching, accelerator choice (GPU/TPU/CPU), and modern LLM serving stacks like vLLM and TGI.
Responsible AI & the Human Side
ML in production isn't only code — it's UX patterns for uncertain predictions, the cross-functional team that ships and maintains a model, and the fairness, privacy, and transparency obligations that come with deploying it.
Anomaly & Outlier Detection
Finding the rare and the unexpected — statistical, distance, density, isolation, and reconstruction methods, the extreme class-imbalance problem, and how to set a threshold.
Privacy & Federated Learning
Training on sensitive data without exposing it — differential privacy and the ε budget, DP-SGD, federated learning, and the fundamental privacy–utility trade-off.
CI/CD, Continuous Training & MLOps Maturity
How ML pipelines automate themselves: Google's MLOps maturity levels 0/1/2, the difference between deploying a model and deploying a training pipeline, and the CI/CD/CT triggers that retrain without humans.
Feature Stores
The system that serves the same feature values to training and inference: offline vs online stores, point-in-time correctness, and how a feature store kills train–serve skew and feature duplication.
Model Registry, Versioning & Governance
The system of record for production models: registries and promotion stages, full lineage and reproducibility, model cards, and the audit/compliance trail that answers 'which model decided this, and why?'.
ML in Practice Quiz
Test your understanding across the whole stack — feature engineering, deployment patterns, monitoring, inference serving, and responsible AI.
Fraud Detection at Scale
Production fraud systems at Nubank, Stripe, and Uber — real-time risk scoring, graph-based detection, adversarial adaptation, and the extreme class-imbalance problem at billion-transaction scale.
Content Moderation at Scale
How platforms classify, rank, and remove harmful content — multiclass toxicity detection, human-in-the-loop labeling, active learning, and the policy enforcement architecture used by Pinterest, Zillow, and DoorDash.
Optimizing LLM Inference
A production case study — constraints in depth (TTFT, ITL, goodput, memory, cost), model-level vs compute-level optimizations, the metric→lever map, and a scoping decision tree for choosing what to optimize first.