AI/ML Services

What Each Service Does

These are API-shaped AI services — you send data, receive structured predictions, and never manage models:

Each exposes a low-level boto3 API and is priced per request or unit of work. The design question is never "how do I host this" — it is whether the accuracy and customization ceiling of a pre-built model fits your task.

Calling the APIs

rekognition.detect_labels(
    Image={"S3Object": {"Bucket": "b", "Name": "p.jpg"}},
    MinConfidence=80,
)
comprehend.detect_sentiment(Text=text, LanguageCode="en")

lex.post_text(botName="OrderBot", userId="u1",
              sessionAttributes={}, inputText="where is my order")

polly.synthesize_speech(Text="<speak>Hello <break/></speak>",
                        OutputFormat="mp3", VoiceId="Joanna",
                        TextType="ssml")

Video work is asynchronous: StartLabelDetection returns a job id you poll with GetLabelDetection. Pass confidence thresholds explicitly — the default may return dozens of low-value labels you would otherwise filter in code.

Choosing and Pitfalls

Per-request pricing wins at modest volume and loses at scale — model the break-even before committing a high-throughput pipeline to managed APIs. When the task needs your domain's terminology, evaluate the custom options (Rekognition Custom Labels, Comprehend custom classification) — they require labeled training data of your own.

Traps: assuming video analysis is synchronous (it is not); shipping user text to Comprehend without a PII/privacy review; forgetting region availability — not every service runs in every region; and treating confidence scores as probabilities rather than ranking signals.

Interview: name the service by input modality and output type, then cite one operational constraint (async jobs, pricing curve, or region limits) that shaped your choice.