Azure AI
Hands-On
A practical, screenshot-driven guide covering Azure AI services end-to-end — from training custom image classifiers and deploying object detection apps to working with Azure OpenAI, DALL·E, and real-time audio models.
About This Guide
What You’ll Build
Image Classification
Train a Custom Vision model to classify cats, dogs, and lions — then retrain with big cat datasets from Kaggle.
Object Detection
Build a fruit detector that draws bounding boxes around objects in images using Azure Custom Vision.
Streamlit Web Apps
Deploy prediction apps on Linux VMs and Streamlit Cloud — with webcam support via GitHub integration.
Computer Vision (OCR)
Set up Azure Computer Vision and build a Flask OCR web app to extract text from uploaded images.
Azure OpenAI + RAG
Deploy GPT-4o in Azure AI Foundry, connect your own documents via AI Search, and publish a chat web app.
DALL·E & Audio
Generate images from text prompts using DALL·E 3, and experiment with real-time audio using GPT-4o audio models.
Lab Contents
Image Classification with Custom Vision
Create resource, upload images, tag and train — achieve 100% precision & recall in Iteration 1.
Classify Big Cats
Download Kaggle big cat dataset (cheetah, leopard, lion, tiger) and train a multiclass classifier.
Prediction via Python API
Publish trained iterations, obtain prediction URL & key, and call the API from Python code.
BigCat Streamlit App on Linux
Install Streamlit on Rocky Linux, build an upload-and-predict app, run on port 8501.
Object Detection — Fruit Detector
Tag bounding boxes for 7 fruit classes (15+ images each), train, publish, and test Quick Test.
Object Detection App + Webcam
Deploy fruit detector app on Streamlit Cloud via GitHub with live webcam and snapshot support.
Computer Vision & OCR Flask App
Create Computer Vision resource, build Flask app on Ubuntu VM, expose on Azure port 5000.
Azure OpenAI Chat Playground
Create Azure OpenAI resource, deploy GPT-4.1, explore Chat Playground in Azure AI Foundry.
RAG — Add Your Own Data
Connect Azure AI Search, upload documents, configure semantic chunking, and query your data.
Deploy OpenAI as a Web App
One-click deploy from AI Foundry to Azure App Service with Entra ID auth and CosmosDB chat history.
Real-Time Audio — GPT-4o
Deploy gpt-4o-realtime-preview in East US 2, enable microphone, and hold live voice conversations.
Image Generation — DALL·E 3
Deploy DALL·E 3 in the Images Playground and generate images from text prompts.
Tech Stack
Key Concepts
Iterative Model Training
Each time you retrain a Custom Vision model, a new iteration is created. Iterations let you compare model versions and roll back — making it safe to add new tags and improve accuracy without losing your previous working model.
Classification vs. Object Detection
Image classification labels the whole image. Object detection goes further — it identifies what objects are present and where they are, returning bounding box coordinates for each detected object.
RAG with Azure AI Search
Retrieval-Augmented Generation grounds GPT responses in your own documents. Azure AI Search indexes your data into chunks; when a user asks a question, the most relevant chunks are retrieved and passed to GPT as context before generating the answer.
Who Is This For
AI-900 / AI-102 Students
Hands-on companion material for Azure AI Fundamentals and AI Engineer certifications.
MCT Trainers
Live lab reference for delivering Azure AI courses to enterprise and government audiences.
AI / ML Developers
Quickly spin up real Vision and OpenAI workloads with working Python code included.
Solutions Architects
Evaluate Azure AI service options for computer vision, NLP, and generative AI use cases.
DevOps Engineers
Learn how to deploy AI-powered apps to Azure App Service and Streamlit Cloud via GitHub.
Cloud Beginners
Step-by-step screenshots make every lab followable with zero prior AI experience required.
Guide Details
Download the Hands-On Guide
12 practical Azure AI labs — Custom Vision, Computer Vision, Azure OpenAI, DALL·E 3, and real-time audio. Free to download, no sign-up required.