HPE Private Cloud AI Solutions Sample Questions:

1. An architect is designing an AI solution to create a medical chatbot that assists doctors by answering questions based on the latest published medical research. The system must be highly reliable, and its answers must be traceable to the source publications. The customer has highlighted that their internal data science team lacks the expertise for complex model retraining but can manage data ingestion pipelines.
Given the customer requirements, which architectural components should the architect include in the solution design? (Select all that apply.)
```
Customer Requirements:
- AI Use Case: Medical Research Q&A Chatbot
- Key Constraint: Information must be current and verifiable.
- Team Skills: Limited AI model training expertise.
- Data Source: Continuously updated database of medical journals.
```

A) A vector database to store embeddings of the medical research papers.
B) A Retrieval-Augmented Generation (RAG) framework to query the vector database.
C) A process for daily fine-tuning of the base LLM on all new research papers.
D) A data ingestion and embedding pipeline to process and add new research to the vector database.
E) A pre-trained, general-purpose Large Language Model (LLM).
F) An edge-optimized server for model deployment in individual hospital rooms.

2. A customer in the public sector wants to use AI to analyze live video feeds from city-wide cameras to automate public safety tasks like detecting traffic accidents or identifying security threats in real-time.
This is their first major AI initiative.
Which AI use case does this scenario represent?

A) Drug Discovery
B) Recommender Systems
C) Computer Vision / Intelligent Video Analytics
D) Natural Language Processing (NLP)

3. An architect is designing an infrastructure solution for an AI workload that involves processing massive datasets for training. The goal is to minimize data transfer latency between the storage system and the GPUs in the compute nodes. The architect wants to enable the NVIDIA GPUs to fetch data directly from the NVMe storage array, bypassing the server's CPU and main memory.
Review the proposed architectural components:
```
- Compute Nodes: HPE ProLiant DL380a Gen11 with NVIDIA H100 GPUs
- Storage: HPE GreenLake for File Storage
- Interconnect: Ethernet with RoCE support
```
Which technology must be enabled and properly configured across these components to achieve this direct GPU-to-storage data path? (Choose 2.)

A) NVLink
B) Multi-Instance GPU (MIG)
C) Confidential Computing
D) GPUDirect Storage (GDS)
E) RDMA over Converged Ethernet (RoCE)

4. A data science team has trained a deep learning model for image classification. While the model achieves 99.8% accuracy on the training dataset, its accuracy drops to only 75% on a new, unseen validation dataset.
The team provides the following training metrics:
```
- Training Epochs: 500
- Training Dataset Size: 1,000 images
- Model Parameters: 15 million
- Training Accuracy: 99.8%
- Validation Accuracy: 75.3%
```
What is the most likely cause of this performance discrepancy?

A) The model is underfitting due to an insufficient number of training epochs.
B) The model is overfitting to the training data and cannot generalize to new data.
C) The learning rate used for training was set too low.
D) The model has too few parameters to learn the features effectively.

5. An architect is presenting to a customer who is an 'Early AI User'. The customer is unsure about the differences between their two main project ideas.
Project 1: Create an internal chatbot that can answer HR policy questions by accessing the live employee handbook stored on a shared drive.
Project 2: Teach a general-purpose chatbot to adopt the company's formal communication style and tone for drafting official press releases.
How should the architect categorize the primary AI task for each project? (Choose 2.)

A) Project 1 is a model training workload.
B) Project 1 is a RAG (Retrieval-Augmented Generation) workload.
C) Project 2 is a fine-tuning workload.
D) Project 2 is a classic inferencing workload with no model customization.

Solutions: