Gen AI Leader
Deep Dive
How a machine learning project moves from raw data to a monitored model in production, and which Google Cloud tools correspond to each stage. Plus, the eight business and technical criteria for choosing the right foundation model among the available alternatives.
Contents
The machine learning lifecycle describes the complete flow that any ML or gen AI project goes through, from raw data to a model reliably running in production. Google Cloud organizes this flow into five stages.
1. Data ingestion
Collecting and bringing raw data — from databases, APIs, logs, sensors, files — into a centralized location where it can be processed.
2. Data preparation
Cleaning, transforming, labeling, and structuring the data so it's usable by a model: removing duplicates, normalizing formats, labeling examples, splitting into training/validation sets.
3. Model training
Adjusting the model's parameters using the prepared data, whether training from scratch, doing fine-tuning on an existing foundation model, or applying lighter techniques such as prompt tuning.
4. Model deployment
Making the trained model available for real applications to consume, typically through an endpoint or API, with the infrastructure needed to serve predictions at scale.
5. Model management
Monitoring the model's performance in production, detecting model drift (performance degradation as real-world data shifts away from the training data), and retraining or updating the model when necessary.
Each stage of the lifecycle has a set of Google Cloud services designed specifically for that task. The exam expects you to be able to match the right stage with the right tool.
1. Ingestion
Storage and streamingCloud Storage, BigQuery, Pub/Sub — centralized storage and streaming of raw data.
2. Preparation
Labeling and transformationAgent Platform data labeling, Dataprep, Dataflow — labeling, cleaning, and transformation of data.
3. Training
Training and tuningTraining on Agent Platform (formerly Vertex AI), Vertex AI Studio (for tuning foundation models), Google Cloud TPUs for intensive training workloads.
4. Deployment
Endpoints and Model GardenAgent Platform endpoints and Model Garden (access to and deployment of pre-trained models).
5. Management
Monitoring and driftAgent Platform Model Monitoring — drift detection, continuous evaluation, and model degradation alerts.
TPU and AI Hypercomputer
TPUs (Tensor Processing Units) are Google's specialized hardware, designed specifically to accelerate ML/gen AI training and inference workloads, and are part of Google Cloud's broader AI Hypercomputer infrastructure.
There's no "best" foundation model in the abstract — the right choice depends on the business use case. Google Cloud identifies eight key criteria to evaluate.
| Criterion | Business question it answers |
|---|---|
| Modality | Does the use case require only text, or also image, audio, or video input/output? |
| Context window | How much information (documents, conversation history) does the model need to process at once? |
| Security | What privacy controls, regulatory compliance, and data protection does the model provider offer? |
| Availability and reliability | Does the model have adequate SLAs, consistent uptime, and support for critical production workloads? |
| Cost | What is the price per token/request, and does it fit the expected usage volume and the project's budget? |
| Performance | How accurate, fast, and low-latency is the model for the specific task that needs to be solved? |
| Fine-tuning | Does the model allow fine-tuning with proprietary data to specialize it for a specific domain or task? |
| Customization | How flexible is the model at adapting to specific needs (grounding, tools, brand tone) beyond fine-tuning? |
Watch out for this
The exam often sets up scenarios where the "most powerful" model isn't the correct answer: if the use case is a simple FAQ chatbot with a limited budget, a smaller, cheaper model with good performance on that specific task is usually the best business choice, not necessarily the largest or most expensive model available.
Did you get this topic?
Test what you just learned
An insurance company needs to process claims that include photos of damage, scanned documents, and text notes from the adjuster, and wants to generate a combined summary in a single response. Which foundation model selection criterion is most decisive for this use case?
Put what you learned to the test with original Gen AI Leader exam questions.
Start Gen AI Leader mock examUnlimited mock exams with the Pro plan
Log in to track your progress.