INTRODUCTION TO GENERATIVE AI

1. Generative AI in Healthcare Diagnostics (Bloom’s Level: Analyze | Company Tag: Google, Meta, Wipro) You are part of an AI team building a Generative AI assistant to help healthcare professionals diagnose medical scans (X-rays/CT/MRI). The system should: • Generate annotated insights from raw scans. • Explain uncertain predictions. • Detect anomalous cases and query for human review. Tasks: a) Compare generative vs. discriminative model choices for generating insights from raw images. Which model is more suitable for content generation versus classification, and why? b) Propose an architecture that integrates: • A CNN/encoder backbone • A diffusion-based generative model for synthetic augmentation • A transformer or LLM component for explanation and narrative output Sketch the high-level architecture with data flow. c) Identify potential ethical risk/failure modes (e.g., hallucination, bias) in such a healthcare system and specify two mitigation strategies you would implement. Why this matters: Modern AI solutions require you to justify architecture choices for generation vs classification, integrate models like transformers and diffusion networks, and handle ethical concerns in high-stake domains. ________________________________________ 2. LLM-Powered Customer Support Chatbot Design (Bloom’s Level: Evaluate & Create | Company Tag: Amazon, Microsoft, Infosys, Oracle) Your task is to design an LLM-based multilingual customer support chatbot for a global e-commerce platform that must: • Process user queries in different languages. • Provide contextual and accurate responses. • Avoid false or harmful outputs (hallucinations). Tasks: a) List three core model design choices (e.g., tokenizer, embeddings, output control) you would make to ensure scalability and multilingual support, and explain the rationale behind each choice. b) Provide a system-level diagram showing components such as: • User interface • Language detection • LLM + prompt logic • Safety / filter module • Logging & monitoring Label how data flows through each component. c) Risk Scenario: If the deployed model hallucinates incorrect policies that mislead users about returns/refunds, propose a technical solution that minimizes hallucination risk while balancing latency & user satisfaction. Why this matters: Tech companies often test your ability to reason about real issues (e.g., safety, scalability) as you design end-to-end solutions and ensure customer-facing products are robust.