The "Buy" Strategy

In the previous lesson, we built a model from scratch using SQL. But sometimes, the smartest engineering decision is not to build.

Google has trained massive models on the world's data (Google Images, Google Translate, etc.). These are exposed as ML APIs.

Exam Rule: If a requirement says "Identify generic objects in images" or "Transcribe audio," and it doesn't specify custom usage (like interpreting X-Ray scans), using an API is always the correct answer. It is cheaper, faster, and more accurate than anything you can build yourself.

---

1. The Core Vision & Language APIs

Vision API

• Capabilities: Object detection (Chair, Cat), Face detection (Smile, Eyes Open), OCR (Text extraction), Explicit Content Detection (SafeSearch), Landmark detection.
• Use Case: An app that lets users upload receipts and extracts the total cost (OCR).

Natural Language API (CNL)

• Capabilities: Sentiment Analysis (Positive/Negative), Entity Extraction (finding "Paris" or "Google" in text), Syntax analysis, Content Classification.
• Use Case: Analyzing customer feedback tweets to see if a product launch was successful.

Translation API

• Capabilities: Detection of language, Translation between 100+ languages.
• Advanced: AutoML Translation allows you to provide a glossary for custom terminology (e.g., specific medical terms).

Speech-to-Text (STT) & Text-to-Speech (TTS)

• STT: Chirp / Universal Speech Model. Handles noisy audio, multiple speakers (diarization).
• Use Case: Transcribing call center logs for analytics.

---

2. Video Intelligence API

Don't confuse this with the Vision API.

• Vision API: Process a single image.
• Video Intelligence API: Process a video file. It adds the dimension of Time.
• Features: Shot change detection, Label detection (e.g., "Dog appears at 00:05 and leaves at 00:10"), Explicit content detection.

---

3. Code Example: Using the Vision API

Notice how simple this is. No training loop. No validation set. Just a request.

from google.cloud import vision

def analyze_image(path):
    client = vision.ImageAnnotatorClient()

    with open(path, "rb") as image_file:
        content = image_file.read()

    image = vision.Image(content=content)

    # Detect features
    response = client.label_detection(image=image)
    labels = response.label_annotations

    print("Labels found:")
    for label in labels:
        print(f"{label.description} (Confidence: {label.score:.2f})")

# Output:
# Cat (Confidence: 0.98)
# Whiskers (Confidence: 0.95)

---

4. Hierarchy of Decision Making

graph TD
    Start{Need Image Analysis} --> Custom{Is the object rare/custom?}
    
    Custom -->|No (e.g., Car, Tree)| API[Use Vision API]
    Custom -->|Yes (e.g., Specific Circuit Board Part)| Data{Do you have lots of data?}
    
    Data -->|No| Labeling[Use Data Labeling Service]
    Data -->|Yes| Train{Expertise Level?}
    
    Train -->|Low| AutoML[Use AutoML Vision]
    Train -->|High| CustomTrain[Use custom TensorFlow/PyTorch]
    
    style API fill:#34A853,stroke:#fff,stroke-width:2px,color:#fff
    style AutoML fill:#F4B400,stroke:#fff,stroke-width:2px,color:#fff
    style CustomTrain fill:#4285F4,stroke:#fff,stroke-width:2px,color:#fff

5. Summary

• APIs are the "Low Code" tier 0 solution.
• No maintenance: Google updates the model. You just call the endpoint.
• Scalability: Handles 1 request or 1 million requests automatically.
• Limit: You cannot change how it thinks. If it thinks a "hotdog" is a "sandwich," you can't fix it. (For that, you need AutoML).

In the next lesson, we bridge the gap. What if you want the ease of an API but need it to recognize your specific products? Enter AutoML.

---

Knowledge Check