Applications of Foundation Models
Free AWS Certified AI Practitioner (AIF-C01) practice — 6 questions on Applications of Foundation Models, with explanations. No sign-up. Full 12-question mixed test →
A financial services company is building a document processing system that must extract structured data from handwritten loan applications, typed forms, and scanned contracts. The system needs to handle tables, checkboxes, signatures, and custom fields unique to their business. They require 99.5% accuracy on their custom fields and need to process 50,000 documents daily. Which approach BEST meets these requirements?
Amazon Textract handles OCR, layout analysis, tables, and forms excellently but cannot be fine-tuned for custom business fields. Fine-tuning a Bedrock FM on the structured text output from Textract allows the model to learn company-specific field patterns and extraction logic while leveraging Textract's proven OCR capabilities. This two-stage approach is the standard architecture for custom document understanding at scale.
A healthcare research organization needs to generate synthetic patient records for ML model training while maintaining statistical properties of real data distributions (age, diagnosis correlations, treatment outcomes). The synthetic data must pass statistical tests for distribution matching but contain zero actual patient information. They have 500,000 real patient records. Which solution provides the MOST appropriate balance of privacy and data utility?
Fine-tuning with differential privacy is the state-of-the-art approach for generating synthetic data that maintains statistical properties while providing mathematical privacy guarantees. The fine-tuned model learns the underlying distributions and correlations in the data, and differential privacy ensures individual patient records cannot be reconstructed. This is the only option that provides both provable privacy and high-fidelity statistical distributions.
An e-commerce company wants to implement a product recommendation system that explains why each product is recommended to increase customer trust. The system must process user browsing history, purchase history, and current session behavior in real-time (under 200ms latency). Which architecture BEST satisfies these requirements?
Amazon Personalize provides real-time recommendations optimized for sub-200ms latency and exposes the reasoning metadata (user-item interactions, item features used). Rule-based explanation templates using this metadata are fast, cost-effective, accurate, and satisfy the trust requirement without adding significant latency. This is a proven pattern for explainable recommendations at scale.
A legal firm needs to redact personally identifiable information (PII) from 100,000 legal documents in 15 languages before sharing them with external auditors. The documents contain standard PII (names, addresses) as well as legal-specific entities like case numbers, client IDs, and attorney-client privileged communications. Redaction accuracy must exceed 99.9% as missed PII creates legal liability. Which solution provides the HIGHEST accuracy?
Amazon Comprehend DetectPiiEntities has proven high accuracy for standard PII, and custom entity recognition can be trained specifically for legal entities with labeled data. The multi-pass approach (standard PII first, then custom entities) with human review of low-confidence predictions is the architecture most likely to exceed 99.9% accuracy. This combines purpose-built tools with domain customization and human oversight.
A manufacturing company wants to implement predictive maintenance by analyzing equipment sensor data (vibration, temperature, pressure) along with maintenance logs, repair manuals, and technician notes. The system should predict failures 48 hours in advance and provide natural language explanations citing specific manual sections. Current data includes 3 years of sensor data (time-series) and 500 maintenance incidents with notes. Which architecture BEST addresses this requirement?
Amazon Lookout for Equipment is purpose-built for predictive maintenance on sensor time-series data and excels at this specific task. RAG with Bedrock allows the system to generate natural language explanations grounded in the actual repair manuals and documentation, providing citations. This separation of concerns (specialized service for prediction, FM for explanation) is architecturally sound and leverages each service's strengths.
A customer service organization is building an AI assistant that must handle 10,000 concurrent conversations across chat, email, and voice channels. The assistant needs to maintain conversation context, access a knowledge base of 50,000 product documents, and escalate to humans when confidence is low. Average conversation length is 8-12 turns. Which architecture provides the MOST cost-effective solution at scale?
Amazon Lex handles conversation management, slot filling, and intent recognition very efficiently and cost-effectively for high-volume scenarios. Using Bedrock only for final response generation (when needed) minimizes expensive FM inference calls. This architecture leverages Lex's optimized conversation engine for the heavy lifting and uses the FM strategically, significantly reducing costs compared to using Bedrock for every turn.
Ready for the real thing?
The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.
$54.99 $17.99 with code FREETEST33 — valid through August 22.
Get my $17.99 deal →