Generative AI Dual-Track Learning Path: From Cloud Platforms to Building Models from Scratch

This project provides a dual-track learning path for generative AI, covering two complementary technical routes: 1) Using Azure AI cloud services (Foundry platform, model fine-tuning, embedding and vector search) to quickly implement applications; 2) Building neural network models from scratch (fully connected networks, CNN, GAN, autoencoders) to deeply understand underlying mechanisms. The dual-track design balances engineering implementation needs and deep learning comprehension, making it suitable for AI learners with different backgrounds.

• Original author/maintainer: srincondelacruz
• Source platform: GitHub
• Original project: GenAI (link: https://github.com/srincondelacruz/GenAI)
• Release date: June 9, 2026

This project is a practical project for the generative AI module of the Master's degree program in Artificial Intelligence and Big Data. It adopts a dual-track learning design, which not only meets the engineering needs of rapid implementation but also provides in-depth learning to understand the internal mechanisms of models.

First Track: Azure AI Cloud Service Practice

• Azure AI Foundry: Unified AI development platform covering the entire process from model selection, fine-tuning to deployment
• Model Fine-tuning: Domain data preparation, training parameter setting, solving overfitting and catastrophic forgetting
• Embedding and Vector Search: Generating high-quality embeddings, building vector databases, implementing semantic search (core of RAG)

Second Track: Building Neural Networks from Scratch

• Fully Connected Networks: Basics of multi-layer perceptrons, understanding forward/backward propagation and gradient descent
• CNN: Convolutional layers, pooling layers, batch normalization, implementing architectures like LeNet/ResNet
• GAN: Variants like DCGAN and conditional GAN, solving mode collapse and training instability
• Autoencoders: Implementation of standard/sparse/variational autoencoders, learning efficient data representations

Dimension	Azure AI Cloud Services	Building Models from Scratch
Learning Curve	Gentle, quick output	Steep, deep understanding
Development Speed	Fast, prototype built in hours	Slow, takes days to tune
Cost Structure	Pay-as-you-go	Mainly invest in computing resources
Interpretability	Black box, relies on API documentation	White box, fully controllable
Application Scenarios	Rapid validation, production deployment	Research innovation, customized needs
Core Skills	Engineering integration, prompt engineering	Mathematical derivation, algorithm optimization

Engineering Practice Capabilities

• Experience in using cloud platform AI services
• Mastery of the complete process of large model fine-tuning
• Ability to build vector databases and semantic retrieval systems
• Proficiency in Keras/PyTorch frameworks

Combination of Theory and Practice

The dual-track design achieves the combination of "knowing how" (tool application) and "knowing why" (underlying principles)

Project Portfolio

• RAG application based on Azure
• Domain-specific fine-tuned model
• GAN image generator implemented from scratch
• Autoencoder dimensionality reduction/denoising demonstration

• Current Status of Large Model Applications: From 2024 to 2025, generative AI moves toward production, and enterprise needs shift to customized models, cost-controllable inference, and interpretable systems
• Rise of Vector Databases: The popularity of RAG architecture promotes vector databases to become infrastructure
• Model Miniaturization Trend: The growth of edge AI and private deployment needs increases the importance of model building and optimization capabilities

Target Audience

• Learners with Python and machine learning basics
• Developers who want to systematically master generative AI
• Students in AI engineering/research directions
• Technical managers

Prerequisite Knowledge

• Python programming
• Linear algebra and calculus
• Machine learning basics (loss functions, gradient descent, etc.)

Learning Suggestions

• Path selection: First complete the Azure cloud service path for quick output, then dive into the self-developed path
• Practice is key: Implement immediately after learning concepts
• Community participation: Record problem solutions and submit improvement suggestions to the original project

This GenAI project provides a structured dual-track learning path, suitable for engineers who want to get started quickly and researchers who pursue depth. Its value lies not only in the comprehensiveness of technical content but also in demonstrating effective learning methods: first gain a global perspective then dive into details, master both tools and principles, and prepare effectively for the rapid iteration of generative AI technology.