Course Content
MODULE 1 Generative AI Foundations
Build a solid conceptual and architectural understanding.
0/7
1.1 Introduction to Generative AI
22:08
1.2 Probability, Tokens & Model Behavior
38:07
1.3 Modern Model Architectures
01:04:30
1.4 The Generative AI Pipeline End-to-End
30:00
1.5 Industry Landscape & Role of Generative AI
32:54
Foundations of Generative AI — Core Concepts Assessment
00:10:00
Build & Analyze a Mini Neural Network From Scratch Using NumPy
MODULE 2 Deep Learning Essentials for Generative AI
This module builds the technical backbone required to understand and create generative AI systems. Learners move from basic neural network mechanics into the advanced architectural concepts that power modern LLMs. The lessons drill into how networks learn, how attention replaces recurrence, and why deep learning techniques like residual connections, layer normalization, and transformer-based patterns dominate current AI systems. The module also covers the practical realities of training—optimizers, loss functions, precision formats, GPU requirements, and distributed strategies—ensuring learners can reason about model performance, stability, and scalability. By the end, students understand the essential engineering foundations behind any production-grade generative model.
0/7
2.1 Neural Network Fundamentals
30:00
2.2 Modern Deep Learning Architecture Patterns
01:19:00
2.3 Attention Mechanism — The Real Engine Behind LLMs
30:00
2.4 Training Dynamics of Deep Networks
30:00
2.5 Deep Learning Infrastructure
01:30:00
Deep Learning Foundations — Neural Networks & Training Dynamics Assessment
00:10:00
Constructing a Neural Network Engine Using Pure NumPy
MODULE 3 Transformer Architecture & LLM Internals
This module provides a complete, practical understanding of the architecture powering modern generative AI systems: the Transformer. The content demystifies how these models process language, how attention mechanisms work, and why Transformers dominate every state-of-the-art model from GPT to Gemini to LLaMA. You’ll learn how the encoder and decoder blocks function, how they differ, and how they interact in more complex tasks like translation and summarization.
MODULE 4 Embeddings, Vector Databases & Semantic Search
A large part of the module focuses on self-attention, the mathematical engine that enables models to reason over long sequences and understand relationships between tokens. You’ll dismantle the Query–Key–Value mechanism, attention scores, softmax scaling, and multi-head attention, gaining a concrete understanding of what is actually computed inside each layer. You will also analyze cross-attention, used heavily in instruction-tuning, multi-modal pipelines, and closed-book question answering.
MODULE 5 Retrieval-Augmented Generation (RAG) Systems
This module focuses on Retrieval-Augmented Generation (RAG), the architecture that turns large language models into reliable, knowledge-grounded systems suitable for enterprise use. You’ll learn why LLMs hallucinate, where their knowledge limits lie, and how retrieval solves these limitations by supplementing the model with external facts at inference time.
Generative AI

Why Transformers Dominated

Transformers won because they solve 3 problems old AI couldn’t:

  1. Parallel processing (RNNs processed sequentially → slow)

  2. Long-context reasoning

  3. Scalability (more data = better performance)

Attention Mechanism (QKV Explained Simply)

You don’t need math; you need intuition:

  • Query (Q): What the current token wants

  • Key (K): What each token offers

  • Value (V): The information attached to each token

Attention = “How much each token should influence the next one?”

Multi-head attention = multiple “perspectives” learning different patterns.

Decoder-Only vs Encoder-Decoder

  • Decoder-only (GPT-style):
    Best for text generation and reasoning.

  • Encoder-decoder (T5-style):
    Best for translation, summarization, and structured tasks.

Modern systems use decoder-only because:

  • simpler

  • cheaper

  • more general-purpose

  • easier to scale

Context Windows

Defines how much the model can “remember” at once.
Larger context windows enable:

  • document reasoning

  • long conversations

  • large RAG chunks

  • multi-step workflows

KV Cache

Critical for speed.
Instead of recomputing everything every token, models store attention states.
This reduces inference cost by up to 80%.

Scaling Laws

More data + more parameters + more compute → predictable performance increases.
This is why companies train gigantic models — the scaling curve rewards them.

Previous
Next
0% Complete