Slide 1: - Pre-trained APIs (Vision, Speech, Natural Language, Video, Tables): Ideal for
Perceptual tasks when you lack Data or ML Skills
- BigQuery ML: Use CREATE MODEL, ML.EVALUATE, ML.PREDICT in SQL; Good if Data
already in BigQuery and Problem fits Predefined models
- AutoML: Graphical Interface on Vertex AI; Handles Data Preprocessing, Model
Selection, and Hyperparameter Tuning for You
- Custom Training: Develop in Vertex AI Workbench or Colab Enterprise; Choose
Frameworks (TensorFlow, PyTorch); Tune Hyperparameters; Deploy via Custom
Containers
Slide 2:
Data Type: Pre-trained APIs Process Tabular, Image, Text, Video, Audio; BigQuery ML
Only Tabular; AutoML & Custom Support All Types
- Data Size: Pre-trained APIs Require None; BigQuery ML & AutoML Need Moderate
Tabular Datasets; Custom Training Best for Large, Diverse Datasets
- Expertise: Pre-trained & AutoML Require Minimal ML or Coding; BigQuery ML Needs
SQL; Custom Training Demands ML and DevOps Skills
- Hyperparameter Tuning: Automatic in BigQuery ML & Custom; Not Available in
Pre-trained APIs or AutoML
- Training Time: Pre-trained is Immediate; AutoML Moderate; BigQuery ML Quick on SQL;
Custom Longest Due to Full Pipeline Setup
- Budget: Custom Training and AutoML Incur Compute and Storage Costs; Pre-trained
APIs Incur Per-call Fees; BigQuery ML Charges Per Query and Storage
Slide 3:
- Align Option with Team Skills, Data Readiness, and Business Urgency
- Use Pre-trained APIs for Chatbots, Image Tagging, Sentiment Analysis Out-of-the-Box
- Use BigQuery ML as a Quick Benchmark Before Investing in Heavier Pipelines
- Use AutoML to Prototype Custom Models Without Code; Transition to Custom Training
When You Outgrow AutoML
- Custom Training is the Path to Production-grade, Specialized Models with MLOps
Integration on Vertex AI
Slide 4:
- Pre-trained APIs let you consume ready-made models (vision, language, speech, etc.)
without building or hosting them
- Analogy: APIs are like electrical sockets—choose the right adapter and plug in, without
knowing wiring details
- Saves time and cost of data collection, labeling, and model maintenance
Slide 5:
- Natural Language API extracts entities, sentiment, syntax, and categories from text
- Vision API recognizes objects, labels, and faces in images; Video Intelligence API adds
motion and action detection
- Document AI automates text extraction and structure parsing for contracts, invoices, ID
documents
- Dialogflow builds chatbots; Contact Center AI adds LLM-powered virtual agents
Slide 6:
Speaker Notes:
- Use Entity Analysis to auto-tag documents, extract key fields, drive metadata for search
- Use Sentiment Analysis on customer feedback, social media, call transcripts to gauge
satisfaction
- Syntax data can improve downstream NLP models by providing linguistic features
- Category Classification helps route support tickets or organize content by subject area
Slide 7:
- Generative APIs allow you to create new content—text, images, code, audio—on
demand
- Embeddings APIs convert text or multimodal inputs into numerical vectors for similarity
search and clustering
- Gemini multimodal can process and generate across text, image, and video in a single
request
- Codey streamlines developer workflows by autocompleting functions, generating
boilerplate, and answering code questions
