From data pipelines through serving, multi-model GPU infrastructure, FinOps for AI, EU AI Act compliance, and incident response. 23 lessons covering the production engineering that turns a working prototype into a reliable system.
Data → train → deploy → monitor → retrain. The full lifecycle and the research-to-production gap.
Design Netflix Recs, ChatGPT Clone, fraud detection, autonomous coding agent. Capacity, latency, cost reasoning.
Batch (Airflow/dbt) vs streaming (Kafka/Flink). Data contracts, schema registries, idempotency, late-arriving data.
Feast / Tecton / Hopsworks. Point-in-time correct joins, training-serving skew, online/offline stores.
W&B / MLflow / DVC. Track every experiment, version models, manage lifecycle stages, ensure reproducibility.
Docker for ML, locked deps, CUDA-friendly base images, multi-stage builds, layer caching.
TorchServe, BentoML, Triton, vLLM, Modal. REST/gRPC/streaming. Batch vs online.
Code CI + data CI + model CI. GitHub Actions, Vertex AI Pipelines, Kubeflow, deploy gates.
Interleaving, holdouts, counterfactual evaluation, gradual rollouts, regression testing, red teaming.
Evidently / Arize / WhyLabs / Fiddler. KS / PSI, latency tracking, cost monitoring, retraining triggers.
Token economics, prompt versioning, eval gates, LLM gateways (Langfuse/Helicone/Portkey), cost management.
GPU selection, model routing, prompt compression (LLMLingua), spot instances, output budgets, cost attribution.
Triton multi-model, vLLM multi-LoRA, MIG/MPS GPU sharing, autoscaling, spot strategy. 3-7× cost savings.
Quantization, pruning, distillation. Make models smaller and faster for edge or cost-constrained serving.
Prompt injection, model theft, adversarial inputs, PII handling. Defenses for the serving layer.
EU AI Act enforcement deadline 2 Aug 2026. NIST AI RMF, automated model cards, data lineage, fairness metrics, high-risk classification.
EU AI Act risk tiers, NIST AI RMF, model cards, algorithmic impact assessments, and sector-specific compliance.
AWS SageMaker, Azure ML, GCP Vertex AI. Compare services, pricing, architecture, when to pick which.
Runbooks for ML incidents (drift, cost spike, prompt injection, infra outage). Rollback in <60s. Chaos drills. Post-mortems.
Capstone — integrate everything. 16-point hardening checklist. Reference architectures for chatbots, fraud, content moderation.
Differential privacy, DP-SGD, federated learning with FedAvg, GDPR compliance, and machine unlearning.
Two AI winters killed careers and companies.
DICOM file format, PACS protocols, HU conversion, photometric interpretation, anonymization for ML, and HIPAA-aligned clinical imaging pipelines.
10 interactive labs — hands-on exercises for this track
You type google.com — trace the journey from your browser through DNS servers to the actual website
10 million users hit your app — watch how traffic is distributed across servers
An API getting hammered — build a rate limiter that protects it without blocking real users
REST, batch, or streaming — the same model deployed three ways has completely different tradeoffs.
Your model was 95% accurate at launch. Three months later: 80%. This is how you catch and fix that.
Live production dashboard for an ML model — latency, throughput, accuracy, and distribution shift. Inject drift and watch multi-signal alerting catch problems.
Split traffic between model v1 and v2, watch conversion, latency, and revenue accumulate. Wait for p<0.05 before promoting a winner.
Mirror production traffic to a new model without ever showing its outputs to users. Compare v1 vs v2 in real time and promote only when safe.
See how a centralized feature store eliminates training/serving skew. Toggle between two-pipeline and feature-store architectures.
5 model versions with promote, rollback, compare, and lineage actions. See why a registry is just 'git for models'.
575 questions across 23 modules — check how well you understood this track.