Task 2.1: Choose appropriate modeling approaches for ML and AI solutions.

• AAmazon Rekognition, detecting text in the scanned images

  Incorrect

  Rekognition is an image and video analysis service. Detecting words in an image does not link a field name to its value or rebuild a table.

• BAmazon Comprehend, detecting entities in the scanned files

  Incorrect

  Comprehend finds entities such as names and dates in text. It does not turn a form layout into key-value pairs and table rows.

• CAmazon Textract, with forms and tables analysis

  Correct

  Textract analysis returns form data as key-value pairs and extracts tables with their cells, which is exactly the structure the scoring model needs from scanned forms.

• DA custom object detection model trained on SageMaker AI

  Incorrect

  A custom model could locate regions on a page, but it would need labeled data, training and extra logic to read and pair the text, which Textract already provides as a managed service.

When a managed AWS AI service already solves the task, it is usually the fastest and cheapest path. Document structure (forms, tables) is Textract's job; images in general are Rekognition's; plain text understanding is Comprehend's.

• ALinear Learner

  Incorrect

  Linear Learner is a supervised algorithm for classification and regression. Without labels of normal and abnormal readings it has nothing to learn from.

• BDeepAR forecasting

  Incorrect

  DeepAR forecasts future values of time series. It predicts what comes next rather than scoring how unusual each existing reading is.

• CRandom Cut Forest

  Correct

  Random Cut Forest (RCF) is the built-in unsupervised algorithm for anomaly detection, such as spotting when a sensor sends abnormal readings, and it needs no labels.

• DObject2Vec

  Incorrect

  Object2Vec learns embeddings of pairs of objects, for example to find similar support tickets. It is not an anomaly detector.

Match the problem type to the algorithm family: unlabeled anomaly scoring points to RCF (or IP Insights for IP addresses), labeled tabular prediction to XGBoost or Linear Learner, and forecasting to DeepAR.

• ASend every email to a large foundation model in Amazon Bedrock with a classification prompt

  Incorrect

  A prompt-based classifier can work, but paying for a large model's input and output tokens on 400,000 emails a day costs far more than a trained classifier, and the labeled data would go unused.

• BTrain an Amazon Comprehend custom classifier on the labeled emails

  Correct

  Comprehend custom classification trains a classifier on your own categories from labeled examples, then classifies documents in real time or in asynchronous jobs. It is managed and built for exactly this kind of routing.

• CWrite a custom PyTorch training script and host the model on a SageMaker AI endpoint

  Incorrect

  This can reach good accuracy, but it means building, tuning and operating the model and its endpoint, which is more effort than the managed service the requirements point to.

• DUse Amazon Comprehend built-in sentiment analysis to route the emails

  Incorrect

  Sentiment returns positive, negative, neutral or mixed. It cannot assign the organization's own 12 categories.

Weigh managed services, custom models and foundation models against volume, labels and effort. With good labels, fixed classes and very high volume, a trained managed classifier is usually cheaper per prediction than a general-purpose foundation model.

• AA vector knowledge base built from nightly CSV exports of the tables

  Incorrect

  Vector retrieval returns the text chunks most similar to the question. It cannot sum and group millions of rows, so totals assembled from a few chunks would be wrong.

• BA GraphRAG knowledge base on Amazon Neptune Analytics

  Incorrect

  GraphRAG connects entities and relationships across document chunks for multi-step reasoning over text. It does not run aggregate queries over warehouse tables.

• CA knowledge base that uses Redshift as a structured data store

  Correct

  A knowledge base connected to a structured data store converts natural-language questions into queries for that store, such as SQL, retrieves the result and generates the answer, so the totals come straight from the tables.

• DFine-tuning a foundation model on a snapshot of the sales tables

  Incorrect

  Fine-tuning teaches patterns from training data, not an exact, current copy of it. The model would produce stale or invented numbers instead of querying the warehouse.

Pick the RAG pattern from the shape of the data: unstructured text uses vector retrieval, connected entities across documents can use GraphRAG, and tables in a database or warehouse use a structured data store that turns questions into queries.

• AReinforcement fine-tuning with a Lambda reward function that runs each query

  Correct

  Reinforcement fine-tuning learns from reward scores instead of labeled input-output pairs, and reward functions can be written in AWS Lambda. It suits tasks whose success can be verified programmatically and where labeled examples are expensive.

• BSupervised fine-tuning on question and ideal-query pairs written by the team

  Incorrect

  Supervised fine-tuning needs labeled examples of the desired output, which is the slow and expensive work the team wants to avoid.

• CDistillation from a larger teacher model

  Incorrect

  Distillation transfers a teacher's ability to a smaller student. It does not use the team's ability to verify each query by running it.

• DA larger context window filled with the full schema

  Incorrect

  Including the schema in the prompt helps, but it does not change the model and does not use the automatic correctness check to improve it.

Supervised fine-tuning learns from labeled pairs, distillation from a teacher's answers, and reinforcement fine-tuning from a reward signal. When correctness can be checked by code but ideal answers are costly to write, rewards are the cheaper signal.

• AThe Priority service tier

  Correct

  Priority delivers the fastest response times for a price premium over standard on-demand pricing, needs no reservation, and is meant for customer-facing workloads that do not warrant 24x7 capacity.

• BThe Reserved service tier

  Incorrect

  Reserved capacity is booked for one or three months with minimum tokens-per-minute commitments. That is the round-the-clock reservation the team does not want.

• CThe Flex service tier

  Incorrect

  Flex gives a discount for workloads that can handle longer processing times. It trades speed for cost, the opposite of what is needed.

• DBatch inference jobs

  Incorrect

  Batch inference processes prompts asynchronously in bulk. It cannot answer a live user in a conversation.

Bedrock offers Reserved, Priority, Standard and Flex tiers, plus batch inference. Choose by how much latency matters and whether capacity is needed all the time.

• AAmazon Polly

  Incorrect

  Polly converts text to speech; it does not understand user intents or manage a conversation.

• BAmazon Comprehend

  Incorrect

  Comprehend analyzes text for entities, sentiment and similar insights; it is not a conversational bot service.

• CAmazon Lex V2

  Correct

  Amazon Lex V2 is an AWS service for building conversational interfaces with voice and text, providing natural language understanding and automatic speech recognition for chatbots.

• DAmazon Translate

  Incorrect

  Translate converts text between languages.

For intent-driven conversational interfaces across voice and text, start with a service built for that purpose before building one from models.

• AAmazon Textract document analysis

  Incorrect

  Textract extracts text and structure from documents; it does not judge whether images are offensive.

• BAmazon Comprehend sentiment analysis of file names

  Incorrect

  File names say little about what an image shows.

• CA SageMaker AI k-means model on pixel values

  Incorrect

  Clustering pixels groups similar images but does not identify inappropriate content, and it means building a model.

• DAmazon Rekognition content moderation

  Correct

  Rekognition moderation APIs detect inappropriate, unwanted or offensive content in images and videos for use cases such as user-generated content.

Managed moderation services provide ready-made detection of unsafe visual content for user-generated media.

• AA tabular algorithm such as XGBoost

  Correct

  AWS maps tabular regression, such as estimating the value of a house, to built-in algorithms like XGBoost, LightGBM, CatBoost and Linear Learner, which learn directly from labeled tabular data.

• BA text-to-image foundation model

  Incorrect

  Image generation does not predict numeric prices.

• CThe k-means clustering algorithm

  Incorrect

  K-means groups similar records; it does not predict a numeric target.

• DA Bedrock knowledge base over the sales table

  Incorrect

  Retrieval returns similar records to a prompt; it does not train a regression model on all 200,000 labeled sales.

Match the problem to the model family: labeled tabular regression is a job for classical supervised algorithms, not generative models.

• ADeepAR

  Incorrect

  DeepAR forecasts time series values.

• BObject2Vec

  Correct

  Object2Vec learns embeddings that preserve the relationship between pairs of objects, which can be used to find nearest neighbors; AWS lists identifying duplicate support tickets as an example.

• CRandom Cut Forest

  Incorrect

  RCF scores anomalies; it does not learn similarity between ticket pairs.

• DIP Insights

  Incorrect

  IP Insights learns associations between IPv4 addresses and entities.

Pairwise similarity problems, such as duplicates or routing by resemblance, fit embedding algorithms such as Object2Vec.

• ALatent Dirichlet Allocation (LDA)

  Incorrect

  LDA discovers topics that are not known in advance; here the 40 categories are predefined and labeled.

• BSequence-to-Sequence algorithm

  Incorrect

  Sequence-to-sequence models produce output sequences, such as translations or summaries, not category labels.

• CBlazingText text classification

  Correct

  AWS maps text classification, assigning predefined categories to documents such as books by discipline, to the BlazingText algorithm and Text Classification - TensorFlow.

• DK-means clustering algorithm

  Incorrect

  Clustering groups similar items without using the known labels.

Predefined labels mean supervised classification; unknown groupings mean topic modeling or clustering.