AI Blueprint Projects for HP AI Studio 🚀

Content

• Overview
• Repository Structure
• Data Science
• Deep Learning
• Generative AI
• NVIDIA GPU Cloud (NGC) Integration
• Troubleshooting
• Contact and Support

---

Overview

This repository hosts a growing collection of AI blueprint projects that run end-to-end using Jupyter notebooks, MLflow deployments, and Streamlit web apps. Each blueprint is designed to be plug-and-play, giving you ready-to-use solutions across key AI domains like data science, machine learning, deep learning, and generative AI. All projects are built to work seamlessly with HP AI Studio.

These blueprints make use of local, open-source models including:

• LLaMA (Meta)
• Qwen (Alibaba)
• BERT (Google)
• LLaMA Nemotron (NVIDIA)

In addition, some examples use hosted models from Hugging Face for added flexibility.

✨ Use cases include:

• Data visualization
• Data analysis
• Image classification
• Super resolution
• LLM fine-tuning
• Multi-modal RAG
• Audio translation
• Agentic RAG
• And much more

We’re actively expanding this collection. If you have ideas or want to suggest a blueprint for HP AI Studio, feel free to open an issue. We’d love to hear from you.

🚀 Get Started with HP AI Studio

To run any of these projects quickly and effortlessly, you can download HP AI Studio for free.

---

Repository Structure

# Root Directory
├── data-science/                          # Projects related to classical machine learning and statistical analysis
│   ├── classification-with-svm/           # SVM-based classification implementation
│   └── data-analysis-with-var/            # Vector AutoRegression analysis workflow
│
├── deep-learning/                         # Deep learning applications using popular frameworks
│   ├── classification-with-keras/         # Image classification using Keras
│   ├── question-answering-with-bert/      # QA system built on top of BERT model
│   ├── recommendation-system-with-tensorflow/  # TensorFlow-based recommendation engine
│   ├── spam-detection-with-nlp/           # NLP-driven spam classifier
│   ├── super-resolution-with-fsrcnn/      # Image enhancement using FSRCNN
│   └── text-generation-with-rnn/          # RNN-based generative model for text
│
├── generative-ai/                         # Generative AI applications across text, code, and image
│   ├── automated-evaluation-with-structured-outputs/  # Eval pipeline for structured generation
│   ├── code-generation-with-langchain/    # Code synthesis using LangChain
│   ├── fine-tuning-with-orpo/             # ORPO-based fine-tuning procedure
│   ├── image-generation-with-stablediffusion/  # StableDiffusion-powered image generation
│   ├── text-generation-with-langchain/    # Text generation leveraging LangChain stack
│   ├── text-summarization-with-langchain/ # Summarization pipeline using LangChain
│   └── vanilla-rag-with-langchain/        # Basic Retrieval-Augmented Generation with LangChain
│
├── ngc-integration/                       # Projects leveraging NVIDIA GPU Cloud and libraries
│   ├── agentic-rag-with-tensorrtllm/      # RAG system using TensorRT-LLM and agentic planning
│   ├── audio-translation-with-nemo/       # Speech translation with NVIDIA NeMo
│   ├── data-analysis-with-cudf/           # RAPIDS cuDF-based data manipulation
│   ├── data-visualization-with-cudf/      # Visualizations using GPU-accelerated tools
│   └── vacation-recommendation-with-bert/ # Recommendation app using BERT embeddings

---

Data Science

The blueprint projects in this folder demonstrate how to build data science applications with HP AI Studio.

We provide 2 blueprint projects, each designed for quick and easy use to help you get started efficiently.

🌸 Classification with SVM

This project is a simple classification experiment focused on predicting species of iris flowers.

It runs on the Data Science Workspace, demonstrating basic supervised learning techniques for multi-class classification tasks.

😷 Data Analysis with VAR

This project explores a regression experiment using mobility data collected during the COVID-19 pandemic.

It highlights how city-level movement patterns changed during the crisis. The experiment runs on the Data Science Workspace.

---

Deep Learning

The blueprint projects in this folder demonstrate how to build deep learning applications with HP AI Studio.

We provide 6 blueprint projects, each designed for quick and easy use to help you get started efficiently.

🖌️ Classification with Keras

This project performs basic image classification using the TensorFlow framework.

It trains a model to classify handwritten digits from the MNIST dataset and runs on the Deep Learning Workspace.

🧠 Question Answering with BERT

This project demonstrates a simple BERT Question Answering (QA) experiment. It provides code to train a BERT-based model, as well as instructions to load a pretrained model from Hugging Face.

The model is deployed using MLflow to expose an inference service capable of answering questions based on input text.

🎬 Recommendation System with TensorFlow

This project builds a simple recommender system for movies using TensorFlow.

It trains on user-item interaction data to predict movie preferences and runs on the Deep Learning Workspace.

🚫 Spam Detection with NLP

This project implements a text classification system to detect spam messages.

It uses deep learning techniques and requires the Deep Learning Workspace for training and inference.

🖼️ Super Resolution with FSRCNN

This project showcases a Computer Vision experiment that applies convolutional neural networks for image super-resolution — enhancing the quality and resolution of input images.

✍️ Text Generation with RNN

This project illustrates how to build a simple character-by-character text generation model.

It trains on a dataset containing Shakespeare's texts, demonstrating the fundamentals of text generation by predicting one character at a time.

---

Generative AI

The blueprint projects in this folder demonstrate how to build generative AI applications with HP AI Studio.

We provide 7 blueprint projects, each designed for quick and easy use to help you get started efficiently.

📊 Automated Evaluation with Structured Outputs

Automated Evaluation with Structured Outputs turns a local Meta-Llama-3 model into an MLflow-served scorer that rates any batch of texts (e.g., project abstracts) against arbitrary rubric criteria.

• Generates scores locally via llama.cpp (no data leaves your machine)
• Registers the evaluator as a pyfunc model in MLflow
• Exposes a REST /invocations endpoint
• Ships two front-ends — a Streamlit dashboard and a pure HTML/JS UI — for instant human-friendly interaction and CSV download.

Code Generation with Langchain

This notebook performs automatic code explanation by extracting code snippets from Jupyter notebooks and generating natural language descriptions using LLMs. It supports contextual enrichment based on adjacent markdown cells, enables configurable prompt templating, and integrates with PromptQuality and Galileo for evaluation and tracking. The pipeline is modular, supports local or hosted model inference, and is compatible with LLaMA, Mistral, and Hugging Face-based models. It also includes GitHub notebook crawling, metadata structuring, and vector store integration for downstream tasks like RAG and semantic search.

Fine Tuning with ORPO

This project demonstrates a full-stack LLM fine-tuning experiment using ORPO (Open-Source Reinforcement Pretraining Objective) to align a base language model with human preference data. It leverages the Z by HP AI Studio Local GenAI environment, and uses models such as LLaMA 3, Gemma 1B, and Mistral 7B as foundations.

We incorporate:

Galileo PromptQuality for evaluating model responses with human-like scorers (e.g., context adherence) TensorBoard for human feedback visualization before fine-tuning A flexible model selector and inference runner architecture A comparative setup to benchmark base vs fine-tuned models on the same prompts

Image Generation with Stable Diffusion

This notebook performs image generation inference using the Stable Diffusion architecture, with support for both standard and DreamBooth fine-tuned models. It loads configuration and secrets from YAML files, enables local or deployed inference execution, and calculates custom image quality metrics, such as entropy and complexity. The pipeline is modular, supports Hugging Face model loading, and integrates with PromptQuality for evaluation and tracking.

Text Generation with LangChain

This notebook implements a full Retrieval-Augmented Generation (RAG) pipeline for automatically generating a scientific presentation script. It integrates paper retrieval from arXiv, text extraction and chunking, embedding generation with HuggingFace, vector storage with ChromaDB, and context-aware generation using LLMs. It also integrates Galileo Prompt Quality for evaluation and logging, and supports multi-source model loading, including local Llama.cpp, HuggingFace-hosted, and HuggingFace-cloud models like Mistral or DeepSeek.

Text Summarization with LangChain

This project demonstrates how to build a semantic chunking and summarization pipeline for texts using LangChain, Sentence Transformers, and Galileo for model evaluation, protection, and observability. It leverages the Z by HP AI Studio Local GenAI image and the Meta Llama 3.1 model with 8B parameters to generate concise and contextually accurate summaries from text data.

Vanilla RAG with LangChain

This project is an AI-powered vanilla RAG (Retrieval-Augmented Generation) chatbot built using LangChain and Galileo for model evaluation, protection, and observability. It leverages the Z by HP AI Studio Local GenAI image and the Meta Llama 3.1 model with 8B parameters to generate contextual and document-grounded answers to user queries about Z by HP AI Studio.

---

NVIDIA GPU Cloud Integration

The blueprint projects in this folder demonstrate how to integrate NVIDIA NGC (NVIDIA GPU Cloud) resources with HP AI Studio.

We provide 5 blueprint projects, each designed for quick and easy use to help you get started efficiently.

🤖 Agentic RAG with TensorRT-LLM

This project contains a single integrated pipeline—Agentic RAG for AI Studio with TRT-LLM and LangGraph—that implements a Retrieval-Augmented Generation (RAG) workflow using:

TensorRT-backed Llama-3.1-Nano (TRT-LLM): for fast, GPU-accelerated inference. LangGraph: to orchestrate an agentic, multi-step decision flow (relevance check, memory lookup, query rewriting, retrieval, answer generation, and memory update). ChromaDB: as a local vector store over Markdown context files (about AI Studio). SimpleKVMemory: a lightweight on-disk key-value store to cache query-answer pairs.

🎙️ Audio Translation with NeMo

This project demonstrates an end-to-end audio translation pipeline using NVIDIA NeMo models. It takes an English audio sample and performs:

1. Speech-to-Text (STT) conversion using Citrinet
2. Text Translation (TT) from English to Spanish using NMT
3. Text-to-Speech (TTS) synthesis in Spanish using FastPitch and HiFiGAN

All steps are GPU-accelerated, and the full workflow is integrated with MLflow for experiment tracking and model registration.

📈 Data Analysis with cuDF

In this project, we provide notebooks to compare the execution time of dataset operations using traditional Pandas (CPU) versus NVIDIA’s cuDF, a GPU-accelerated drop-in replacement for Pandas. This example is presented in two different formats:

• Original Example Notebook: This version, created by NVIDIA, runs the entire evaluation within a single notebook. It includes downloading the data and restarting the kernel to activate the cuDF extension.

• Data Analysis Notebooks: These notebooks use preprocessed datasets of varying sizes from the datafabric folder in AI Studio. The evaluation is split across two notebooks—one using Pandas (CPU) and the other using cuDF (GPU)—with performance metrics logged to MLflow.

📡 Data Visualization with cuDF

This project is a GPU-accelerated, interactive exploratory data analysis (EDA) dashboard for the OpenCellID dataset. It uses Panel and cuDF to deliver lightning-fast geospatial analysis and visualization.

You can explore cell tower distributions by radio type, operator, country, and time window — rendered live on an interactive map with full GPU acceleration.

🌍 Vacation Recommendation with BERT

This project implements an AI-powered recommendation agent that delivers personalized travel suggestions based on user queries.

It leverages the NVIDIA NeMo Framework and BERT embeddings to understand user intent and generate highly relevant, tailored vacation recommendations.

---

Troubleshooting

This section provides solutions for common issues users may encounter when working with AI Blueprint projects in HP AI Studio:

1. Check Hardware Compatibility Each project’s README includes recommended minimum hardware specifications (e.g., RAM, VRAM). Make sure your system meets these requirements—especially when working with large models or during deployment, as insufficient resources can cause failures.

2. Models or Datasets Not Visible After Download If you download models or datasets while your workspace is running, they might not appear in the workspace. In such cases, restart your workspace to ensure they are properly recognized.

3. Connection or SSL Errors in Notebooks If you encounter SSL certificate or connection errors while accessing websites from notebooks (especially on restricted networks), verify your network settings. Consider using a proxy to bypass restrictive network constraints.

4. File or Path Not Found Errors Ensure that all required files and directories are correctly placed as specified in the project’s README. If any paths or files are missing, create or move them to the correct locations.

5. GPU Not Available For projects requiring NVIDIA GPUs, verify GPU availability by running nvidia-smi in the terminal. Ensure that a compatible GPU is accessible and has sufficient free memory to run the project.

6. Deployment Errors Despite Meeting Requirements Even if your hardware meets the specs, limited available RAM or VRAM can cause deployment issues. Close other running workspaces or programs to free up memory.

7. API Timeout Issues API requests triggered for the local MLflow deployment in the AI Studio have a response timeout limit (approximately 4 minutes). For long-running tasks or large inputs, use the provided notebooks instead of the API to avoid timeout errors.

---

Contact and Support

• Troubleshooting: Refer to the Troubleshooting section of the main README in our public AI-Blueprints GitHub repo for solutions to common issues.

• Issues & Bugs: Open a new issue in our AI-Blueprints GitHub repo.

• Docs: AI Studio Documentation.

• Community: Join the HP AI Creator Community for questions and help.

---

Built with ❤️ using HP AI Studio.