Prepare and test your skills

Prepare and test your skills

The Vertex AI Model Development Lifecycle is the end-to-end workflow for building a custom machine learning model on Google Cloud, from preparing data to serving predictions. Vertex AI (opens in a new tab) provides managed infrastructure, tracking, and deployment services at each stage, so teams can experiment quickly without building their own platform. Data moves forward from one managed service to the next, which lowers complexity and lets a business create a model tuned to its own needs.
The lifecycle follows a fixed order:
Data preparation in Vertex AI is the process of turning raw data into clean inputs that a model can learn from. A feature is a characteristic or attribute of an entity that is used as input to train a model or to make inferences, and feature engineering is the process of transforming raw data into those features. A managed dataset is a dataset object created in and hosted by Vertex AI, and the training set is typically the largest portion at about 80%, separate from the validation set used during training and the test set used after training.
Vertex AI Feature Store (opens in a new tab) is the managed service for storing, serving, and managing machine learning features. Feature data sources in BigQuery are registered through feature groups and features, and feature views aggregate features from one or more sources into an online store instance for serving. Feature serving then exports or fetches values for the next stage, with online serving retrieving the latest values for real-time inference and offline or batch serving exporting high volumes of data, including historical data, for model training.
Training in Vertex AI is the process where a model learns patterns and relationships from the training set so it can make inferences on new data. Training can run in pre-built containers for frameworks such as TensorFlow and PyTorch, or as custom training jobs, without the team managing servers and accelerators. Because the platform provisions compute and handles scaling, failure recovery, and optimization, teams can start new runs quickly and focus on code and data.
Vertex AI Experiments tracks each experiment run across the workflow. An experiment run records:
Vertex AI TensorBoard is a managed, scalable service that visualizes those experiments using the open-source TensorBoard interface. Data flows from Vertex AI Training to TensorBoard for persistent storage, where data scientists can debug training, track performance metrics, and compare runs collaboratively. Vertex ML Metadata and ML pipelines add lineage and automation, since ML pipelines are portable and scalable container-based workflows that can use prebuilt Google Cloud pipeline components to define repeatable steps.
Tuning and evaluation in Vertex AI is the stage where model settings are improved and model quality is measured before deployment. Vertex AI Vizier is a black-box optimization service for tuning hyperparameters and other parameters across repeated experiment runs. This tuning loop depends on experiment tracking, because each trial's parameters flow into Vizier and the resulting metrics flow back into Experiments and TensorBoard for comparison.
Vertex AI Model Evaluation is a managed service that assesses the performance and quality of machine learning models. The service generates evaluation metrics and visualizations, such as the F1 score that balances precision and recall for tasks like information extraction, so teams can find weaknesses and decide whether to deploy or improve the model. Explainable AI adds interpretation by providing tools to understand inferences, including insights into feature importance and model behavior.
Deployment in Vertex AI is the process of placing a trained model into production so it can be used for inference on unseen data. Vertex AI Model Registry (opens in a new tab) is the central repository where teams manage the lifecycle of models, organize versions, track new training, and deploy a chosen version to an endpoint directly or by using aliases. Storing the model in the registry before deployment creates a clear boundary between development, where versions change often, and production, where a stable version serves traffic.
Vertex AI Inference is the service that runs new data points through the deployed model to calculate an output, which is also called operationalizing a model or putting it into production. Inference requests flow from applications to the endpoint, and the endpoint returns predictions from the deployed model version.
| Inference mode | Input pattern | Result pattern |
|---|---|---|
| Batch inference | A group of inference requests processed together | Results are output together in one file |
| Online inference | Individual data points sent as they arrive | Results are returned in real time |
Exam tip: Choose batch inference when predictions can be grouped into one file, and choose online inference when a single new data point needs a real-time prediction.
Custom models built with Vertex AI (opens in a new tab) are machine learning models trained or adapted on an organization's own data to solve its specific problems. A generic pretrained solution handles common tasks the same way for everyone, while a custom model learns the patterns, language, and customer behavior unique to one business. This difference creates strategic value because the custom model can power products, services, and operations that competitors using only generic models cannot copy.
Custom models that fit unique business needs are models that learn from business-specific features and examples. In machine learning, a feature is a characteristic or attribute used as input to train a model or to make inferences, and feature engineering is the process of transforming raw data into those features. Vertex AI Feature Store is the managed service for storing, serving, and managing those features, with online serving that retrieves the latest values for real-time inferences and offline serving that exports high volumes of historical data for training. Because the workflow depends on the organization's own feature data, the resulting model reflects its customers, products, and processes.
Vertex AI supports four data types for custom work:
A managed dataset hosted by Vertex AI is typically split so the training set is the largest portion, typically 80%, where the model learns patterns, while separate validation and test sets check performance during and after training. A true positive, such as correctly predicting that a specific customer will purchase a jacket, shows how a custom model is judged on business-relevant outcomes, and metrics such as F1 score measure accuracy when both precision and recall matter.
The Vertex AI custom model lifecycle is the ordered flow where data moves from preparation through deployment and each stage depends on the previous stage. Vertex AI Model Registry is the central repository where teams manage that lifecycle, organize models, track versions, and deploy versions over time. The managed flow lets teams move from idea to production without managing the underlying platform.
The lifecycle moves in order:
Inference from a deployed custom model takes two forms. Batch inference runs a group of requests together and outputs the results in one file, while online inference returns real-time inferences on individual data points. Choose batch inference when the business can wait for grouped results, such as nightly scoring, and choose online inference when the user experience depends on an immediate answer.
Differentiated user experiences are experiences where a model understands an organization's content and responds in its voice and context. A foundation model (FM) (opens in a new tab) is a model trained on broad data that can be adapted to a wide range of downstream tasks, and a Transformer is the neural network architecture behind most state-of-the-art generative models where an encoder converts input text into an intermediate representation and a decoder converts it into output using self-attention. Training a Transformer from scratch requires significant resources, while fine-tuning a pretrained Transformer for a specific application is more efficient. Model distillation extends this efficiency by letting a smaller student model learn to mimic the output of a larger teacher model for lower-cost serving.
Vertex AI connects custom knowledge to generative experiences through grounding and retrieval. A vector is a numerical representation of text, images, or video that captures relationships between inputs, and applications use embeddings to recognize complex meanings and semantic relationships. Information retrieval finds and retrieves relevant documents from a large collection based on a user's query, and retrieval-augmented generation grounds model responses in real-time data from that retrieval. Data flows from the organization's corpus to an index, from the index to the retrieval step at prompt time, and from the retrieval step into the model prompt, so answers reflect current business information.
The choice of generative capability depends on the experience being built. Gemini handles text and multimodal prompts that include images or video, Imagen generates images, Codey generates code, and text embeddings support search, classification, or clustering. Vertex AI Agent Engine (opens in a new tab) then provides the services to deploy, manage, and scale AI agents in production, so the custom experience can grow without the team managing the underlying agent infrastructure.
Generic pretrained solutions are pre-built models for common tasks that accelerate time to value without training. Examples include Optical Character Recognition that extracts text from images and files, Speech-to-Text that converts spoken language into written text, and Vertex AI Translation that translates text between languages. Vertex AI Model Garden is the place to manage AI models procured from Google Cloud Marketplace, which also favors reuse over building from scratch. Choose a generic solution when the need is standard, the data is not proprietary, and speed matters more than uniqueness.
Three-column comparison of Generic Pretrained API, Custom Model on Vertex AI, and Fine-tuned Foundation Model across data used, best fit, and strategic effect.
| Option | Data used | Best fit | Strategic effect |
|---|---|---|---|
| Generic pretrained API | Broad public data | Common tasks such as OCR, transcription, and translation | Fast to launch but the same for every competitor |
| Custom model on Vertex AI | Organization's own image, text, tabular, and video data | Unique terminology, proprietary behavior, and specialized predictions | Harder to build but creates a capability competitors do not have |
| Fine-tuned foundation model | Broad base plus organization-specific examples | Adapting Gemini or an embedding model to company voice and knowledge | Balances strong base quality with business-specific differentiation |
Exam tip: Choose a generic pretrained API when the task is common across businesses, and choose a custom or fine-tuned Vertex AI model when advantage comes from learning unique business data and delivering a distinct user experience.
Vertex AI (opens in a new tab) is a machine learning platform that lets organizations train and deploy machine learning models and AI applications. Operational integration means connecting those models to products, workflows, and processes so they produce business value. Business differentiation comes from using unique data to train or tune models, choosing deployment options that balance cost, performance, and scalability, and embedding model outputs into daily operations.
Operational integration is the work of making a custom model part of a real product or process rather than a standalone experiment. Vertex AI provides services such as:
Applications can call these services through APIs and SDKs (opens in a new tab) available for languages such as Python, Go, Java, and Node.js. In a GDC Sandbox environment, Vertex AI uses the regular Google Cloud version of the API, requires a Google Cloud billing account, and setup includes creating or identifying a project, enabling billing, enabling the Vertex AI API, authenticating, and installing a client library.
A concrete integration example is Google Security Operations (opens in a new tab) with Vertex AI. This integration supports automated malicious email analysis, contextual entity enrichment, flexible playbook orchestration, and custom data transformation. It works with any generally available model accessible through the Vertex AI API endpoint, including foundation models, third-party models, open-source models, fine-tuned models, and custom-trained models. Generally available models are recommended for production because Preview models may have instability, breaking changes, or limited support. The integration provides actions such as Analyze EML, Describe Entity, Execute Prompt, Ping, and Transform Data. To set up the integration, an organization creates a service account, grants Identity and Access Management roles, and chooses an authentication method: Workload Identity, which is recommended and uses short-lived tokens through service account impersonation, or a service account JSON key, which is a static, long-lived secret key file.
Operationalizing a generative AI model follows a lifecycle:
Evaluation loops through model selection, prompt engineering, tuning, and optimization. Not all models require deployment: Google's foundation models already have endpoints that are specific to a Google Cloud project and are immediately available. However, if a foundation model is tuned, it must be deployed to an endpoint.
Exam tip: Google foundation models already have project-specific endpoints, but a tuned model must be deployed to an endpoint before use.
Cost, performance, and scalability determine whether an integrated model can support long-term business differentiation. Deploying a model associates an endpoint and physical machine resources with the model for serving online, low-latency predictions. Organizations choose between a fully managed environment and a self-managed environment.
| Deployment option | Control | Management | Example |
|---|---|---|---|
| Fully managed | Select machine type and accelerator type | Vertex AI instantiates and manages the resources | Deploy a model to an endpoint |
| Self-managed | More fine-grained control over resources | You manage the resources yourself | Serve models on Google Kubernetes Engine (GKE) |
When choosing resources, organizations balance anticipated traffic, latency requirements, and budget. If lower cost is a priority, they might tolerate higher latency with lower-cost machines, as in a test environment. Model size also matters: size is typically measured by the number of parameters. Larger models can learn more complex patterns and produce higher-quality responses, but larger models in the same family can have higher latency and costs. Cost is related to capabilities and parameter count, and models can be metered differently: some charge by input and output tokens, while others charge by node hours used while the model is deployed. Not all models support features such as tuning and distillation.
In a Google Distributed Cloud Sandbox environment, Vertex AI can run in two usage models. A CPU-based workload runs in the Sandbox user cluster without GPUs and is slower. A GPU-based workload uses GPU support in the Sandbox AI Optimized SKU and configures the workload to use GPUs associated with the sandbox-gpu-project project; that project exists only in Sandbox instances provisioned with the AI Optimized SKU. On Cloud Marketplace, some AI models have regionalized pricing that changes based on the location of the endpoint receiving requests. Organizations using multiple regions should verify that requests go to an endpoint in their preferred location. Eligible third-party AI models can also draw against an organization's Flexible Savings Plans.
Custom model training uses an organization's own data to tailor AI solutions to specific needs. Model tuning improves performance and helps a model adhere to specific output requirements for specialized tasks. To tune a model, an organization builds a training dataset and selects a tuning method, such as:
The dataset size and tuning method depend on the model and what is being optimized; specialized, niche tasks can require a smaller dataset to achieve significant improvements.
Prompt design is often the first customization step. If a model's responses work well, further customization may not be needed. If the model underperforms or hallucinates, additional techniques can help. Function calling and Vertex AI Extensions expand model capabilities. Function calling is asynchronous and does not require credentials in code. Vertex AI Extensions provide prebuilt options for complex tasks, but because extensions return and call functions for the model, they require credentials in code. Grounding anchors model responses to verifiable sources to reduce hallucinations. Retrieval augmented generation is a grounding technique that uses search to find relevant information and adds it to the model prompt; retrieval search uses vector embeddings and vector databases.
Model evaluation helps assess how prompts and customizations affect performance. Metrics-based evaluations can be automated and scaled quickly, but they can oversimplify results and miss context. Automatic side-by-side evaluation compares the output of two models against ground truth, with a third model selecting the higher-quality response. It is quicker than human evaluation and available on demand, but it requires a larger model than the ones being evaluated and can have inherent biases, so human evaluations are still needed. Evaluation datasets contain prompt and ground truth pairs, and they should include diverse examples that align with the task. Responsible AI also requires testing models for safe and responsible use; a model is only as good as the data used, so inaccurate or incomplete data can lead to suboptimal performance or false outputs.
Choosing the right AI approach affects business impact because it determines how quickly a solution reaches the market and how well it fits the problem. For classification or detection use cases, an organization first checks whether a pre-trained traditional AI model can meet its requirements. Pre-trained traditional models include Document AI, Vision AI, Natural Language API, and Video Intelligence API. If a pre-trained model meets the requirements, it should be used. If not, the organization checks whether enough training data is available.
A decision tree starting from a business use case: generative tasks like summarization go straight to generative AI, while classification or detection tasks first try a pre-trained traditional AI model, then branch on training data availability and on control versus go-to-market priority to choose custom traditional training, generative AI with tuning, or generative AI with prompt engineering.
When sufficient training data exists, the decision depends on priorities. If high control of model training is required — such as using a preferred algorithm, custom loss functions, explainability features, specific layer counts, learning rates, or other hyperparameters — custom training of a traditional AI model is appropriate. If faster go-to-market is the priority, generative AI is appropriate; specialized use cases can be improved with model tuning, such as supervised tuning for classification, sentiment analysis, or entity extraction. If a training dataset is not available or is not large enough to custom train a model, generative AI models with prompt engineering can be used, and they can be tuned further with data examples.
Traditional AI is particularly effective for predictive use cases with structured data, such as classification, regression, time series forecasting, and clustering. Generative AI fits summarization, content generation, advanced transcription, and use cases that process multiple modalities like text, images, videos, or audio. The two approaches can also combine: output from a traditional AI model can become part of the prompt for a generative AI model. Examples include:
Stakeholders beyond data scientists and ML engineers influence model selection:
Business differentiation from custom models comes from using unique data and specialized tuning to create tailored products and services. Vertex AI supports this through custom model training, pre-built models and algorithms, and MLOps capabilities for model deployment, monitoring, and management. Pre-built models accelerate time to value, while custom training tailors solutions to specific needs. MLOps streamlines machine learning workflows so models can be deployed, monitored, and managed consistently.
For organizations with strict data sovereignty, security, and privacy requirements, Vertex AI on Google Distributed Cloud air-gapped (opens in a new tab) can run services entirely within a data center. This air-gapped deployment keeps data inside the organization's infrastructure while providing a familiar Vertex AI experience with the same tools, APIs, and models from Vertex AI on Google Cloud. The benefits include enhanced security and privacy, increased agility, accelerated time to value, and streamlined MLOps. These capabilities support operational innovation and long-term differentiation when organizations integrate models into products, processes, and security workflows, and when they continuously evaluate cost, performance, and scalability as usage grows.
The Vertex AI Model Development Lifecycle runs in order from data preparation to training and experiment tracking to tuning and evaluation to deployment from the Vertex AI Model Registry to an endpoint for inference. Vertex AI provides managed infrastructure, tracking, and deployment services at each stage so teams can experiment quickly without building their own platform.
Batch inference processes a group of inference requests together and outputs the results together in one file, while online inference sends individual data points as they arrive and returns results in real time. Choose batch inference when predictions can be grouped into one file and choose online inference when a single new data point needs a real-time prediction.
Choose a generic pretrained API when the need is standard, the data is not proprietary, and speed matters more than uniqueness, such as Optical Character Recognition, Speech-to-Text, and Vertex AI Translation. Choose a custom or fine-tuned Vertex AI model when advantage comes from learning unique business image, text, tabular, and video data and delivering a distinct user experience.
Untuned Google foundation models already have project-specific endpoints and are immediately available, whereas a tuned foundation model must be deployed to an endpoint before use. After deployment, Vertex AI Inference runs new data points through the deployed model version to calculate outputs.
A specialized electronics manufacturer wants to gain a competitive advantage by automating real-time defect detection on its assembly line. The company has a following set of operational circumstances and technical requirements:
Which Vertex AI approach should the company use to address its unique business needs and differentiate its quality control process?
Last updated