AWS managed AI services

Pre-trained ML services: no model building, token-based pricing, multi-AZ redundancy, specialized hardware behind the scenes. The exam wants you matching use case to service.

Details worth keeping

Comprehend custom classification builds your own document categories, trained from tagged samples in S3, real-time (single doc) or async batch. Custom entity recognition does the same for business-specific terms (policy numbers, escalation phrases), while plain NER extracts the generic entities (people, places, dates).

Transcribe accuracy boosters: custom vocabularies (specific words: brand names, acronyms, with pronunciation hints) plus custom language models (domain context, trained on your text). Use both for best results. It also does voice-based toxicity detection (tone + text cues).

Polly extras: lexicons (expand AWS to “Amazon Web Services”), SSML markup for pronunciation and pauses, several voice engines, speech marks for lip sync and word highlighting.

Rekognition custom labels teach it your logo or products with a few hundred labeled images. Content moderation integrates with A2I to cut human review to 1-5% of volume, and custom moderation adaptors tune it with your labeled images.

Personalize recipes are pre-built algorithms per use case: USER_PERSONALIZATION, PERSONALIZED_RANKING, POPULAR_ITEMS (Trending-Now), RELATED_ITEMS (Similar-Items), PERSONALIZED_ACTIONS (Next-Best-Action), USER_SEGMENTATION. Recipes = recommendations.

Kendra learns from user feedback (incremental learning) and allows manual result tuning.

The A2I flow: high-confidence predictions return straight to the app; low-confidence ones go to human reviewers (your staff, AWS contractors, or Mechanical Turk); reviewed data feeds back into training.

Healthcare variants (HIPAA territory)

AWS hardware for AI

GPU instance families: the P and G series.

Trainium is the AWS training chip (Trn1 has 16 accelerators, around a 50% training cost cut). Inferentia is the inference chip (Inf1/Inf2, up to 4x throughput at roughly 70% lower cost).

Trn and Inf also carry the lowest environmental footprint pitch.