Building a machine learning solution on Google Cloud usually starts with choosing among three options: pre-trained AI APIs, AutoML, and custom training. That trade-off comes up regularly on the Professional Machine Learning Engineer exam because scenario questions hinge on matching a workload's customization needs, timeline, cost tolerance, and team expertise to the right level of abstraction. Making that choice also requires knowing what Google Cloud's pre-trained APIs already provide out of the box so you do not recommend training a model when an API call already solves the task.
Choosing between pre-trained APIs, AutoML, and custom training
The three approaches sit along a clear trade-off between ease of development and control:
- Pre-trained AI APIs: Built for common, general-purpose ML tasks such as image labeling, text analysis, speech transcription, or translation. They work immediately through an API call with no model training required, though they are not a fit when you need domain-specific customization or control over model internals.
- AutoML: A middle ground on customization for teams that want to build their own models without writing code. It fits workloads where minimizing preprocessing and reducing development time matter more than controlling model internals, though it can come at a higher cost even as it simplifies development.
- Custom training: Built for teams with ML expertise that need full control over both the training data and model internals, including custom model architectures. It is the right path for complex or novel problems.
Text and document analysis: Cloud Natural Language API and Document AI
The Cloud Natural Language API analyzes unstructured text for general use cases without domain-specific customization. It provides three core capabilities out of the box:
- Sentiment analysis: Assigns a sentiment score to text. For a sentence such as "John loves his new iPhone that he bought in New York last week," it returns a strongly positive score such as +0.85, driven heavily by the word "loves."
- Entity extraction: Identifies people, places, products, and dates, recognizing John as a person, iPhone as a consumer good, New York as a location, and last week as a date. Each entity includes metadata such as type and salience, and can link to external resources such as Wikipedia when relevant.
- Syntax parsing: Returns each token's part of speech (noun, verb, adjective) and syntactic role (subject, object, or root), such as labeling John as a proper noun and subject and loves as a verb and root. This helps extract relationships and understand how ideas connect within a sentence.
Together, these Cloud Natural Language API capabilities provide a fast way to break down raw text for summarizing content, tagging entities, or building basic understanding into text-heavy workflows.
Document AI classifies, extracts, and processes information from documents automatically using machine learning across four capabilities:
- Document classification: Categorizes documents by type, such as invoices, contracts, claim forms, or tax documents, replacing manual sorting.
- Entity extraction: Pulls key information from unstructured documents, including invoice numbers, dates, vendor names, line items, and totals from invoices, or parties, dates, key clauses, and obligations from contracts.
- Document processing at scale: Processes batches of invoices, contracts, forms, and claims in high volume rather than one document at a time.
- Custom extractors: Pre-built processors handle common document types, while company-specific or industry-specific forms with unique layouts can use a custom extraction solution trained on labeled examples of your own documents.
Document AI also integrates directly with BigQuery through remote connections. BigQuery uses that remote connection to execute Document AI custom extractors natively with SQL commands, allowing extraction inside a query rather than building a separate pipeline to move files back and forth.
Image and video analysis: Cloud Vision API and Video Intelligence API
Cloud Vision API and Video Intelligence API offer similar pre-trained vision features for different inputs: Vision API analyzes images, while Video Intelligence API analyzes videos frame by frame. Their capabilities include:
- Location detection (Cloud Vision API only): Identifies famous landmarks such as St. Basil's Cathedral, returning the name, geographic coordinates, and a matching map reference.
- Image and video labeling: Identifies high-level concepts or objects such as people, streets, or modes of transport for categorizing content or adding metadata.
- Object localization (object detection): Draws bounding boxes around detected objects to show where each appears, supporting counting or tracking specific instances.
- Text detection (OCR): Recognizes and outlines printed and handwritten text for scanning documents, reading signage, or structuring visual data.
- Face detection: Both APIs detect faces, identify facial landmarks such as eyes, nose, and mouth, and estimate emotions including joy, sorrow, anger, and surprise.
On the Professional Machine Learning Engineer exam, keep face detection separate from facial recognition. Neither Cloud Vision API nor Video Intelligence API identifies or verifies a person's identity. They detect faces and facial attributes, not who a face belongs to.
Language and audio: Cloud Translation, Speech-to-Text, and Text-to-Speech
The Cloud Translation API delivers fast, scalable translation with no training or tuning. It translates text across more than 100 languages, detects the source language automatically when it is unknown, and translates full documents such as PDFs or Word files while retaining their original structure and formatting.
Speech-to-Text API transcribes spoken audio into text for captions, voice commands, or call center logs, while Text-to-Speech API synthesizes human-like audio from text with customizable voice, pitch, and speed for virtual assistants or reading apps. Both APIs let you select which model handles a request, and Chirp is one of Google's model families available on both sides.
Several Speech-to-Text details matter for exam scenarios:
- Speaker labeling: Identifies which parts of a conversation came from speaker 1, speaker 2, and additional speakers.
- Real-time and batch modes: Supports both live streaming audio and pre-recorded files.
- Synchronous vs. asynchronous batch recognition: Audio shorter than 60 seconds uses synchronous recognition and returns the transcribed text in the same response. Audio longer than 60 seconds uses asynchronous recognition, starting a long-running operation that you check once processing completes (an hour-long meeting must use the asynchronous path).
- Audio sample rates: Capture audio at 16,000 Hz (16 kHz) or higher so speech frequencies are recorded cleanly. Do not resample existing lower-rate audio upward: resampling 8 kHz audio to 16 kHz will not recover detail that was never recorded, so send it at its native rate.
Personalized e-commerce recommendations: Recommendations AI API
The Recommendations AI API helps e-commerce sites deliver personalized product recommendations to improve engagement and revenue using two strategies:
- Other products you might like: Shows visually or behaviorally similar items to raise click-through rates and engagement, though not necessarily purchase size.
- Frequently bought together: Focuses on upselling complementary items to increase cart size and total revenue.
Applying the decision framework
On the Professional Machine Learning Engineer exam, we would lean toward a pre-trained API whenever a scenario describes a standard text, document, vision, translation, speech, or e-commerce recommendation task that does not require domain-specific customization or control over model internals. When a team needs custom models on its own data with no training code, minimal preprocessing, and faster turnaround, AutoML provides the middle path at potentially higher cost. Custom training belongs where an experienced ML team needs full control over training data, model internals, and novel problem designs.
Our Professional Machine Learning Engineer course covers choosing between pre-trained AI APIs, AutoML, and custom training alongside Gemini Enterprise Agent Platform training workflows and model deployment, with practice questions that drill these distinctions.