-
Translate business challenges into ML use cases
-
Choose the optimal solution (ML vs non-ML, custom vs pre-packaged)
-
Define how the model output should solve the business problem
-
Identify data sources (available vs ideal)
-
Define ML problems (problem type, outcome of predictions, input and output formats)
-
Define business success criteria (alignment of ML metrics, key results)
-
Identify risks to ML solutions (assess business impact, ML solution readiness, data readiness)
-
Design reliable, scalable, and available ML solutions
-
Choose appropriate ML services and components
-
Design data exploration/analysis, feature engineering, logging/management, automation, orchestration, monitoring, and serving strategies
-
Evaluate Google Cloud hardware options (CPU, GPU, TPU, edge devices)
-
Design architectures that comply with security concerns across sectors
-
Explore data (visualization, statistical fundamentals, data quality, data constraints)
-
Build data pipelines (organize and optimize datasets, handle missing data and outliers, prevent data leakage)
-
Create input features (ensure data pre-processing consistency, encode structured data, manage feature selection, handle class imbalance, use transformations)
-
Build models (choose framework, interpretability, transfer learning, data augmentation, semi-supervised learning, manage overfitting/underfitting)
-
Train models (ingest various file types, manage training environments, tune hyperparameters, track training metrics)
-
Test models (conduct unit tests, compare model performance, leverage Vertex AI for model explainability)
-
Scale model training and serving (distribute training, scale prediction service)
-
Design and implement training pipelines (identify components, manage orchestration framework, devise hybrid or multicloud strategies, use TFX components)
-
Implement serving pipelines (manage serving options, test for target performance, configure schedules)
-
Track and audit metadata (organize and track experiments, manage model/dataset versioning, understand model/dataset lineage)
-
Monitor and troubleshoot ML solutions (measure performance, log strategies, establish continuous evaluation metrics)
-
Tune performance for training and serving in production (optimize input pipeline, employ simplification techniques)
