Skip to content

Skill 1.1.1: Extract data from data sources

Question 1 - Foundation Model Selection for Real-Time Customer Service

A financial services company is building a real-time customer service chatbot for their trading platform. The chatbot must process customer queries about account balances, recent transactions, and market orders with sub-second latency. The company's compliance team requires complete data privacy (no data leaves the AWS environment) and the solution must support multi-turn conversations with context retention. The engineering team has budget constraints and wants to minimize ongoing inference costs.

Which foundation model approach should the ML engineer recommend? (Select ONE)

  • A) Deploy a large multimodal model (e.g., 100B+ parameters) on Amazon Bedrock with default on-demand pricing to maximize response quality.
  • B) Use Amazon Bedrock with a smaller, task-specialized model (e.g., Anthropic Claude Haiku or Amazon Titan Text Lite) configured with conversation history in the inference parameters.
  • C) Fine-tune a large foundation model on proprietary financial data using Amazon SageMaker and deploy it to a real-time endpoint with auto-scaling.
  • D) Build a custom deep learning model using Transformer architecture in Amazon SageMaker for complete control over latency and inference costs.
  • E) Use Amazon Bedrock Guardrails without any foundation model, relying solely on rule-based response templates for compliance.
Click to reveal Answer & Explanation

Correct Answer: B

Correct Option Analysis

  • B is Correct: For real-time conversational AI with sub-second latency requirements and cost constraints, selecting a smaller, purpose-built foundation model through Amazon Bedrock (such as Claude Haiku or Titan Text Lite) provides the optimal balance. These models offer fast inference, lower per-token costs, and built-in support for multi-turn conversations through the conversationHistory parameter in the Bedrock API. Data privacy is maintained as Bedrock operates within AWS and does not use customer data to train the underlying FMs.

Distractor Analysis

  • A is Incorrect: Large multimodal models with 100B+ parameters have higher latency (typically multiple seconds for complex queries) and significantly higher per-token costs, violating both the sub-second latency requirement and budget constraints for a high-volume customer service chatbot.
  • C is Incorrect: Fine-tuning a large model on proprietary data introduces unnecessary complexity, cost, and time when a task-specialized FM can handle general customer service queries. Fine-tuning is only needed for domain-specific behavior that prompting cannot address.
  • D is Incorrect: Building a custom transformer model from scratch requires massive datasets, GPU infrastructure, and ongoing operational overhead. This violates the efficiency and budget constraints when Amazon Bedrock already provides built-in conversational capabilities.
  • E is Incorrect: Amazon Bedrock Guardrails is a feature for content filtering and safety controls around an FM—it is not a standalone conversational service. Using rule-based templates alone would not handle natural language queries with the flexibility required for trading platform customer service.

Key Takeaway: For low-latency, cost-effective conversational AI, select smaller task-specialized FMs on Bedrock rather than maximizing model size—latency and cost constraints often outweigh marginal quality gains.

Question 2 - Prompt Engineering vs Fine-Tuning Decision

A healthcare organization is building a clinical documentation assistant using Amazon Bedrock. The assistant must generate medical summaries using specialized healthcare terminology and ICD-10 coding conventions. The evaluation team has already implemented detailed system prompts with three few-shot examples showing the expected terminology and formatting. However, the model still inconsistently applies internal medical abbreviations and occasionally generates summaries in the wrong structure.

The ML engineer notices that with additional few-shot examples (8+), the model improves but still fails ~15% of test cases. What is the MOST appropriate next step? (Select ONE)

  • A) Immediately fine-tune the foundation model using Amazon Bedrock's custom model feature with a dataset of labeled prompt-response pairs.
  • B) Continue expanding the few-shot examples indefinitely until the failure rate drops below 1%, without any fine-tuning.
  • C) Switch to a larger foundation model with more parameters, assuming it will naturally handle medical terminology better.
  • D) Remove all few-shot examples and rely solely on the system prompt to prevent overfitting to specific examples.
  • E) Evaluate whether fine-tuning with supervised learning on curated medical summaries would cost-effectively close the persistent gap.
Click to reveal Answer & Explanation

Correct Answer: E

Correct Option Analysis

  • E is Correct: The progression from prompting to fine-tuning follows a specific order: prompt engineering first, evaluate, then fine-tune only when a consistent gap persists. With a 15% failure rate after thorough prompt engineering efforts, supervised fine-tuning (SFT) on labeled medical summaries is the appropriate next step. Amazon Bedrock supports custom model creation with labeled JSONL data, allowing the model to internalize specialized terminology and structure.

Distractor Analysis

  • A is Incorrect: While fine-tuning is the eventual solution, the answer suggests "immediately" fine-tuning without proper dataset preparation and evaluation. The correct approach involves first preparing quality training data and evaluating whether fine-tuning is justified—which option E captures.
  • B is Incorrect: Adding increasingly more few-shot examples leads to diminishing returns and exceeds the prompt context window limits. This approach also increases token costs per request and does not fundamentally change the model's internal behavior.
  • C is Incorrect: Larger models do not automatically understand specialized medical terminology better. Model scale improves general reasoning but without domain-specific training or robust prompts, the terminology gap will persist—while also increasing inference costs and latency.
  • D is Incorrect: Removing few-shot examples would degrade performance further. The examples are helping (reducing errors from baseline), and removing them would eliminate valuable in-context learning signals.

Key Takeaway: Follow the progression: Prompt (Instructions + Few-Shot) → Evaluate → Fine-Tune only when a measurable gap persists after exhausting prompt improvements.

Question 3 - RAG Architecture for Dynamic Knowledge Base

A legal services firm is building a document question-answering system that must provide accurate answers about contracts, case law, and regulatory filings stored across multiple Amazon S3 buckets. The content is updated daily, and lawyers must be able to cite specific sections of original documents for each answer. The firm has a 3-person engineering team and needs to minimize operational overhead while ensuring answers reflect the most current information.

Which architecture pattern BEST satisfies these requirements? (Select ONE)

  • A) Fine-tune a foundation model monthly on the entire legal corpus using Amazon Bedrock custom models, with a manual versioning system for document updates.
  • B) Use Amazon Bedrock Knowledge Bases connected to S3 data sources, with automatic syncing to ingest document changes and provide source attribution through retrieval.
  • C) Build a custom RAG pipeline using Amazon SageMaker, OpenSearch Serverless, and AWS Lambda functions for complete control over the retrieval process.
  • D) Embed the legal documents directly into the foundation model's system prompt using a large context window (e.g., 200K tokens) for each query.
  • E) Use Amazon Comprehend for entity recognition on legal documents and generate summaries without a foundation model.
Click to reveal Answer & Explanation

Correct Answer: B

Correct Option Analysis

  • B is Correct: Amazon Bedrock Knowledge Bases is the fully managed RAG solution that directly addresses all constraints:
    1. Daily updates: Auto-sync from S3 ensures the vector index is refreshed without manual intervention.
    2. Source citations: Retrieval-augmented responses include source metadata (document, page, chunk) from the vector store.
    3. Minimal operational overhead: Bedrock manages document parsing, chunking, embedding generation, and vector store management.
    4. Accuracy: The model retrieves current passages at query-time rather than relying on stale weights.

Distractor Analysis

  • A is Incorrect: Monthly fine-tuning cannot keep pace with daily content updates. Fine-tuning also cannot provide source citations for specific document sections and requires retraining (hours/days) any time content changes.
  • C is Incorrect: Building a custom RAG pipeline with Lambda, SageMaker, and OpenSearch Serverless provides more control but significantly higher operational overhead. With a 3-person engineering team, managing custom embedding pipelines, chunking logic, and vector index maintenance violates minimal-overhead requirements.
  • D is Incorrect: Embedding entire documents in the prompt is impractical for a large legal corpus (thousands of documents). Context windows have token limits, and per-query costs would be exorbitant. This approach also doesn't provide source citations.
  • E is Incorrect: Amazon Comprehend performs NLP tasks like entity recognition but cannot answer complex questions about legal documents. It provides no retrieval or generative capabilities needed for document Q&A.

Key Takeaway: For dynamic content updates with citation requirements, choose managed RAG (Bedrock Knowledge Bases) over fine-tuning—retrieval keeps knowledge external and immediately refreshable.

Question 4 - Evaluating FM Performance Criteria

A media streaming company is selecting a foundation model for a content recommendation explanation feature. The model must: (1) generate natural-language explanations for why a movie was recommended, (2) process 500 requests per second at peak hours, (3) maintain p99 latency under 2 seconds, and (4) keep monthly inference costs under $10,000. The candidate models are evaluated on these AWS benchmark metrics:

  • Model X: 80% quality score, 300 tokens/sec throughput, $0.008/1K tokens, p99 latency 3.5s
  • Model Y: 75% quality score, 1,200 tokens/sec throughput, $0.003/1K tokens, p99 latency 1.2s
  • Model Z: 85% quality score, 150 tokens/sec throughput, $0.015/1K tokens, p99 latency 4.8s

Which model selection BEST satisfies the performance criteria, and what additional AWS services should be used? (Select TWO)

  • A) Select Model X for its superior quality score above the 80% threshold.
  • B) Select Model Y for meeting latency and cost constraints while maintaining acceptable quality.
  • C) Use Amazon Bedrock's on-demand pricing for all models to simplify cost management.
  • D) Use Amazon Bedrock with Provisioned Throughput to ensure consistent performance for production traffic.
  • E) Select Model Z because quality is the most critical criterion regardless of cost and latency.
Click to reveal Answer & Explanation

Correct Answers: B and D

Correct Options Analysis

  • B is Correct: Model Y meets all hard constraints:
    • Latency: p99 = 1.2s, under the 2-second requirement ✓
    • Throughput: 1,200 tokens/sec supports high request volume ✓
    • Cost: $0.003/1K tokens is the lowest cost, likely within $10K/month ✓
    • Quality: 75% is acceptable ("quality" is 5% below Model X but the hard constraints matter more)
  • D is Correct: Amazon Bedrock Provisioned Throughput is the correct purchase mode for production workloads with strict throughput and latency SLAs. It provides guaranteed capacity with consistent p99 latency, unlike on-demand which may suffer throttling during peak load.

Distractor Analysis

  • A is Incorrect: Model X fails the latency constraint (3.5s p99 > 2s requirement) and has lower throughput (300 tokens/sec) that cannot support 500 requests per second. Quality cannot compensate for viable production constraints.
  • C is Incorrect: On-demand pricing on Bedrock does not guarantee consistent throughput or latency during demand spikes. For 500 requests/sec with p99 latency under 2 seconds, Provisioned Throughput is required to maintain consistent performance.
  • E is Incorrect: Model Z has the highest quality but violates every hard constraint: latency (4.8s > 2s), throughput (150 tokens/sec), and cost ($0.015/1K tokens). Pure quality cannot override feasibility requirements.

Key Takeaway: Model selection must evaluate all constraints (latency, throughput, cost) as hard requirements—not just quality scores. Use Provisioned Throughput for production Bedrock workloads with strict SLAs.

Question 5 - AI Service Selection for Document Processing

An insurance company processes 50,000 claim documents daily, including scanned forms, handwritten medical records, and typed policy documents. The processing pipeline needs to: (1) extract structured data (policy numbers, claim amounts, dates), (2) detect fraudulent claim patterns through entity relationships, and (3) transcribe voicemails from claimants. The engineering team needs to implement this within 2 weeks with minimal ML expertise.

Which AWS service architecture BEST accomplishes these requirements? (Select ONE)

  • A) Fine-tune a foundation model on insurance documents using Amazon Bedrock to handle all three tasks in a single model.
  • B) Use Amazon Textract for document extraction, Amazon Comprehend for fraud pattern detection through entity analysis, and Amazon Transcribe for voicemail conversions.
  • C) Build custom deep learning models in Amazon SageMaker for each task, using specialized architectures (CNNs for documents, RNNs for fraud, Seq2Seq for audio).
  • D) Use Amazon Rekognition for document processing, Amazon Comprehend for transcription, and Amazon Transcribe for fraud detection.
  • E) Deploy a single multimodal foundation model on Amazon SageMaker that processes documents, text, and audio simultaneously.
Click to reveal Answer & Explanation

Correct Answer: B

Correct Option Analysis

  • B is Correct: This solution leverages managed AWS AI services for three distinct, pre-defined tasks:
    • Amazon Textract: Pre-trained for extracting structured text, forms, and tables from scanned documents and images, including handwritten content.
    • Amazon Comprehend: Detects entities, relationships, and key phrases that can support fraud pattern detection through entity co-occurrence analysis.
    • Amazon Transcribe: Purpose-built for converting voicemail audio into text for downstream processing.
    • All three services can be implemented via API calls within the 2-week timeline with minimal ML expertise.

Distractor Analysis

  • A is Incorrect: Fine-tuning a single FM on insurance documents is time-consuming, expensive, and suboptimal for all three tasks. Foundation models are poor at structured data extraction from scanned documents and cannot process raw audio directly. This would take months, not 2 weeks.
  • C is Incorrect: Building custom deep learning models from scratch requires ML expertise, GPU infrastructure, labeled training data, and 3-6 months of development. This contradicts all constraints (2-week timeline, minimal ML expertise).
  • D is Incorrect: Service mappings are wrong: Rekognition is for visual content (not document text extraction), Comprehend cannot transcribe audio, and Transcribe doesn't detect fraud. This would not work at all.
  • E is Incorrect: Multimodal FMs cannot process raw audio or extract structured form data from scanned images reliably. They are also not designed for high-volume (50K/day) document processing at cost-effective rates.

Key Takeaway: For standard, pre-defined tasks (OCR, NLP analysis, speech-to-text), use managed AWS AI services before considering custom ML or foundation models—they save weeks of development and operational overhead.

Question 6 - RAG vs Fine-tuning for Compliance Documentation

A pharmaceutical company needs an internal assistant that can answer questions about drug testing procedures, safety protocols, and regulatory submissions. The assistant must always provide source citations to specific sections of Standard Operating Procedures (SOPs), and any updated SOP must be reflected in answers within 1 hour of the document change. The SOP library contains 5,000 documents that are updated weekly.

Which TWO architectural choices BEST satisfy these requirements? (Select TWO)

  • A) Implement RAG architecture using Amazon Bedrock Knowledge Bases for document retrieval and response generation.
  • B) Fine-tune a foundation model on the 5,000 SOP documents to internalize the knowledge.
  • C) Set up an S3 event notification that triggers automatic sync of Bedrock Knowledge Bases when SOP documents are updated.
  • D) Store all SOP content in the model's system prompt using the latest context window expansion features.
  • E) Use a larger foundation model (200B+ parameters) with a comprehensive system prompt containing company policies.
Click to reveal Answer & Explanation

Correct Answers: A and C

Correct Options Analysis

  • A is Correct: RAG architecture with Bedrock Knowledge Bases is the correct approach because it:
    • Retrieves content dynamically at query time, ensuring answers reflect current document versions.
    • Provides source citations through chunk-level metadata (document title, section, page).
    • Separates knowledge from model weights, allowing immediate updates.
  • C is Correct: Automating document sync through S3 event notifications + Knowledge Bases auto-sync triggers ensures the vector index updates within minutes of a document change. This satisfies the 1-hour reflection requirement without manual intervention.

Distractor Analysis

  • B is Incorrect: Fine-tuning on 5,000 SOP documents embeds knowledge in model weights, but:
    • Updates require complete retraining (hours/days, not 1 hour).
    • No source citation capability for specific SOP sections.
    • Catastrophic forgetting risk when fine-tuning on narrow domain data.
  • D is Incorrect: Embedding 5,000 documents in a system prompt exceeds context window limits, incurs massive per-query token costs, and does not provide source citations. This approach is neither feasible nor cost-effective.
  • E is Incorrect: Larger foundation models do not solve the citation and freshness problems. They cannot hold 5,000 documents in context, cannot cite specific sections, and do not automatically update when documents change.

Key Takeaway: RAG (not fine-tuning) is the correct architecture when requirements include source citations and rapid content update propagation—retrieval keeps knowledge fresh and traceable.

Question 7 - Latency vs. Model Performance Trade-off

A financial trading firm needs to build a real-time sentiment analysis system for news feeds that triggers automated trading decisions. The system must process 10,000 news articles per second, with p99 latency under 50 milliseconds, and operate 24/7 with minimal maintenance. The team is comparing a deep learning model (95% accuracy, 120ms p99 latency) with a traditional ML model (85% accuracy, 15ms p99 latency).

Which solution BEST satisfies the requirements with appropriate performance-accuracy trade-off? (Select ONE)

  • A) Deploy the deep learning model on Amazon SageMaker with multi-model endpoints to improve latency through parallel processing.
  • B) Deploy the traditional ML model on Amazon SageMaker with a real-time endpoint, accepting the accuracy reduction for latency compliance.
  • C) Use a hybrid approach: route high-confidence articles to the traditional ML model and low-confidence articles to the deep learning model through Amazon API Gateway and AWS Lambda.
  • D) Deploy the deep learning model on Amazon Bedrock with custom optimization for latency reduction.
  • E) Increase the deep learning model's batch size to achieve higher throughput, accepting slightly reduced accuracy for the performance gain.
Click to reveal Answer & Explanation

Correct Answer: B

Correct Option Analysis

  • B is Correct: The traditional ML model meets the hard latency constraint (15ms p99 < 50ms requirement) and is deployed correctly on Amazon SageMaker. For algorithmic trading where every millisecond matters, the 85% accuracy might be sufficient if the trading strategy accounts for uncertainty. The hard constraint is latency, not accuracy, making this the only viable option.

Distractor Analysis

  • A is Incorrect: Multi-model endpoints improve resource utilization across multiple models but do not reduce the deep learning model's inherent 120ms p99 latency. The deep model fundamentally violates the 50ms hard constraint regardless of deployment architecture.
  • C is Incorrect: A hybrid routing approach still sends some articles to the deep learning model that takes 120ms, violating the p99 latency requirement. The routing logic also introduces additional latency overhead, making this approach counterproductive for real-time trading.
  • D is Incorrect: Amazon Bedrock hosts foundation models, not custom deep learning models for sentiment analysis. Bedrock also introduces API overhead and cannot achieve 50ms p99 latency for custom models. The service choice is incorrect for this use case.
  • E is Incorrect: Increasing batch size for the deep learning model reduces per-item processing cost but does not reduce latency. Batch processing adds queuing delay, making latency worse. SageMaker batch transform or increased batch size in inference contradict real-time latency needs.

Key Takeaway: When requirements include strict latency SLAs, those constraints become non-negotiable—an 85% accuracy model that meets 50ms p99 is superior to a 95% accuracy model that takes 120ms p99.

Question 8 - Foundation Model Selection Criteria

A global e-commerce platform needs to implement product description generation for 10 million products across 25 product categories. The system must: (1) generate descriptions in 15 languages, (2) process this as a monthly batch job (not real-time), (3) maintain brand voice consistency across all categories, and (4) minimize total infrastructure cost. The team has 6 months to implement the solution.

Which approach BEST balances quality, cost, and operational efficiency? (Select ONE)

  • A) Fine-tune a single foundation model on product descriptions for all 25 categories and 15 languages, then run batch inference monthly through Amazon Bedrock.
  • B) Use Amazon Bedrock on-demand pricing with the largest available foundation model to generate all 10 million product descriptions without fine-tuning.
  • C) Build a custom sequence-to-sequence model in Amazon SageMaker for each product category and language pair (375 models total).
  • D) Use Amazon Polly with neural TTS voices to generate spoken product descriptions for accessibility compliance.
  • E) Deploy a pre-trained model from Amazon SageMaker JumpStart and use SageMaker batch transform for monthly description generation.
Click to reveal Answer & Explanation

Correct Answer: E

Correct Option Analysis

  • E is Correct: Amazon SageMaker JumpStart provides pre-trained sequence-to-sequence models suitable for text generation tasks. Using SageMaker batch transform enables cost-effective monthly batch processing without maintaining always-on endpoints. This approach:
    • Leverages pre-trained capabilities (no full model training from scratch).
    • Capitalizes on batch inference pricing (pay only for compute used).
    • Provides flexibility to customize prompts for different categories and languages.
    • Fits the 6-month timeline with existing solution templates.

Distractor Analysis

  • A is Incorrect: Fine-tuning a single FM on 25 categories and 15 languages requires massive, high-quality labeled data (millions of language-category pairs) and expensive GPU training. A single fine-tuned model would struggle with brand consistency across all combinations and introduce unnecessary complexity.
  • B is Incorrect: On-demand pricing for 10 million descriptions with the largest FM would be prohibitively expensive (10M × ~500 tokens × 15 languages = 75 billion tokens × premium pricing). The largest FM is also overkill for basic product descriptions and introduces unnecessary inference costs.
  • C is Incorrect: Building 375 custom models (25 categories × 15 languages) is an operational nightmare requiring 375 training jobs, 375 endpoints, and significant ML expertise. This contradicts cost minimization and 6-month timeline constraints.
  • D is Incorrect: Amazon Polly converts text to speech; it does not generate product descriptions. This is a fundamental service mismatch.

Key Takeaway: For batch processing at scale, use pre-trained models (JumpStart) with batch transform to optimize cost—avoid fine-tuning unless a specific quality gap exists, and avoid on-demand pricing for massive batch workloads.

Question 9 - Prompt Engineering Best Practices

A customer service organization is implementing a foundation model via Amazon Bedrock to classify support tickets into categories and assign priority levels. Initial evaluation shows that the model does not consistently distinguish between "billing issue" and "technical bug" categories when both involve payment failures. The ML engineer is considering how best to improve classification accuracy without customizing the model weights.

Which TWO approaches should the engineer implement to improve classification without fine-tuning? (Select TWO)

  • A) Provide clear category definitions and inclusion/exclusion criteria in the system prompt.
  • B) Fine-tune the model on the historical ticket dataset to learn the distinction.
  • C) Add few-shot examples in the prompt showing examples of tickets correctly classified as "billing issue" vs "technical bug".
  • D) Increase the temperature parameter to 1.0 so the model explores more classification possibilities.
  • E) Use a larger foundation model to improve classification accuracy through more parameters.
Click to reveal Answer & Explanation

Correct Answers: A and C

Correct Options Analysis

  • A is Correct: Providing explicit category definitions with inclusion/exclusion criteria directly addresses the confusion. For example: "Billing issue: customer was charged incorrectly or payment failed due to card decline. Technical bug: system error prevented transaction completion." Clear boundaries help the model distinguish between adjacent categories.
  • C is Correct: Few-shot examples are the most effective prompt-engineering technique for classification tasks. Providing 3-5 examples of both "billing issue" and "technical bug" tickets (especially edge cases involving payment failures) gives the model concrete patterns to follow without modifying model weights.

Distractor Analysis

  • B is Incorrect: Fine-tuning on historical tickets is the "fine-tuning first" anti-pattern. The question specifically asks for improvements "without customizing the model weights." Fine-tuning should only be considered after prompt engineering (options A and C) has been exhausted and evaluation gaps persist.
  • D is Incorrect: Increasing temperature to 1.0 increases randomness in model outputs, making classification more inconsistent and less deterministic. For classification tasks, lower temperature (0.0-0.3) is preferred to achieve stable, consistent predictions.
  • E is Incorrect: Switching to a larger foundation model is a costly leap that does not specifically address the category confusion problem. Larger models have more general knowledge but do not inherently know the company's specific definitions of billing vs. technical issues without proper prompting.

Key Takeaway: For classification tasks, improve performance through: (1) explicit category definitions in system prompts, (2) few-shot examples of edge cases, and (3) low temperature settings for deterministic outputs—never jump to fine-tuning first.

Question 10 - Interpretability Requirements for Financial Predictions

A credit union is building a loan approval ML model that must predict default probability based on 15 features (income, credit score, employment duration, debt-to-income ratio, etc.). The risk management team requires complete model interpretability—they need to explain every prediction to regulators and understand exactly how each feature contributes to the decision. The model must also update weekly with new loan data and maintain prediction latency under 100ms.

Which modeling approach BEST satisfies the interpretability requirement? (Select ONE)

  • A) Use a deep neural network trained in Amazon SageMaker with declarative interpretability features like SHAP or LIME for post-hoc explanations.
  • B) Use a transparent tree-based model (e.g., logistic regression or decision tree) trained with a SageMaker built-in algorithm, providing inherent interpretability.
  • C) Use a foundation model on Amazon Bedrock with a prompt that includes all 15 features and requests a detailed explanation.
  • D) Use an ensemble of gradient-boosted trees with hundreds of estimators for maximum predictive accuracy.
  • E) Use Amazon SageMaker Autopilot to automatically select the best model without considering interpretability constraints.
Click to reveal Answer & Explanation

Correct Answer: B

Correct Option Analysis

  • B is Correct: For regulatory compliance where every prediction must be explainable, a transparent tree-based model or logistic regression provides inherent interpretability. Logistic regression coefficients directly quantify each feature's contribution (e.g., "each 10-point increase in credit score reduces default probability by 0.8%"). Decision trees provide rule-based decision paths that are auditable and explainable to regulators. SageMaker built-in algorithms (e.g., XGBoost with limited depth, Linear Learner) support this approach with minimal infrastructure overhead.

Distractor Analysis

  • A is Incorrect: Deep neural networks are inherently "black box" models. While SHAP and LIME provide post-hoc explanations, they are approximations that regulators often find insufficient for high-stakes loan decisions. Post-hoc explainers introduce additional complexity and potential inaccuracies in the interpretability layer.
  • C is Incorrect: Foundation models are designed for natural language tasks, not tabular credit risk modeling. They cannot reliably process 15 numerical features for precise probability estimation, lack feature-level interpretability, and add unnecessary cost and complexity to a traditional ML problem.
  • D is Incorrect: An ensemble of hundreds of trees is a "black box" model that violates interpretability. While ensembles may achieve slightly higher accuracy, the risk management team cannot explain individual prediction paths through hundreds of decision trees.
  • E is Incorrect: SageMaker Autopilot optimizes for predictive accuracy by default and may select complex neural architectures that violate interpretability. Autopilot offers limited model transparency and does not prioritize regulatory compliance constraints.

Key Takeaway: When interpretability is a hard regulatory constraint, choose inherently interpretable models (logistic regression, shallow decision trees) over post-hoc explainability—regulators want transparent decision logic, not approximate explanations.