The Complete Agentic AI Engineering Masterclass (2026)

This lecture examines the contrast between Traditional Engineering and Generative AI Engineering, focusing on how their underlying principles, tools, and outcomes differ. Traditional engineering is rule-based and deterministic, relying on fixed logic, databases, frameworks, and testing to build predictable applications such as calculators, blogs, and media platforms. In contrast, generative AI engineering is data-driven and probabilistic, leveraging prompts, embeddings, vector databases, and large language models (LLMs) to produce context-dependent, conditional outputs.

We will analyze how these differences manifest across key aspects:

• Approach: Deterministic rules vs. probabilistic, data-driven reasoning

• Patterns: CRUD/API/DevOps vs. data, models, embeddings

• Rules: Hard-coded logic vs. prompt engineering and parameter tuning

• Outputs: Predictable/fixed vs. probabilistic/conditional

• Tools: Frameworks, databases, Postman vs. LLMs, HuggingFace, vector DBs, evals

• Domains: Classical apps and directories vs. NLP, AI assistants, classification, and prediction tasks

By the end of this lecture, learners will clearly understand how Gen-AI Engineering extends beyond traditional software practices, redefining workflows, toolchains, and solution spaces.

This lecture unpacks the inner workings of Large Language Models (LLMs), breaking down the key components that enable them to understand and generate human-like text. We begin with tokenization, where input text is split into smaller units (tokens) for efficient processing. These tokens are then transformed into embeddings, high-dimensional numerical vectors that capture semantic meaning and relationships between words.

Next, we explore the self-attention mechanism, the core innovation behind transformers, which allows models to weigh the importance of different tokens in a sequence and capture contextual relationships across entire texts. Finally, we examine the prediction process, where the model generates outputs step by step by selecting the most probable next token based on learned patterns from its training data.

By the end of this lecture, learners will understand the step-by-step pipeline of LLMs—from raw text input to meaningful predictions—and how these components work together to power state-of-the-art AI applications.

This lecture focuses on Retrieval-Augmented Generation (RAG), a powerful technique that combines information retrieval with generative AI to produce accurate, context-aware outputs. We begin by explaining the RAG workflow: a user query is transformed into embeddings, relevant documents are retrieved from a vector database, and the retrieved context is passed to a language model to ground its response in reliable knowledge.

We then explore the open-source ecosystem that enables RAG, including:

• Vector Databases (e.g., FAISS, Milvus, Pinecone, Weaviate) for efficient semantic search.

• Embedding Models (e.g., OpenAI, HuggingFace, SentenceTransformers) to convert text into dense vector representations.

• Frameworks and Orchestration Tools (e.g., LangChain, LlamaIndex, Haystack) for building RAG pipelines and integrating retrieval with LLMs.

By the end of this lecture, learners will understand how RAG enhances the reliability of LLMs, the key steps in building a RAG pipeline, and the diverse open-source frameworks available to implement it effectively.

This lecture explores the decision-making process between Fine-tuning and Retrieval-Augmented Generation (RAG) for adapting Large Language Models to specific tasks. We begin by explaining fine-tuning, where a model’s weights are updated on domain-specific data to specialize it for tasks such as classification, sentiment analysis, or domain-restricted text generation. Fine-tuning is best suited for structured, repetitive, or highly specialized tasks where knowledge does not change frequently.

In contrast, RAG is introduced as a more flexible approach that augments a base model with external, retrievable knowledge. RAG is most effective when working with dynamic, evolving information or large knowledge bases, as it avoids retraining and enables models to stay current.

Through real-world examples, we illustrate:

• Fine-tuning for tasks like legal document classification, medical report tagging, or customer intent detection.

• RAG for applications such as enterprise search, question answering, or chatbots requiring access to large and constantly changing datasets.

By the end of this lecture, learners will clearly understand the trade-offs between fine-tuning and RAG, and how to choose the right strategy for different problem domains.

Understand how to set up and run Large Language Models locally, including installation, configuration, and hardware considerations for efficient deployment.

Learn how to interact with locally deployed LLMs through user interfaces and APIs, enabling experimentation, customization, and integration into real-world workflows.

This section explores the open-source ecosystem that empowers developers, researchers, and enterprises to build end-to-end AI solutions without being locked into proprietary systems. We examine the role of platforms that provide pretrained models, libraries for fine-tuning and inference, and toolkits for orchestrating workflows. Key focus areas include:

• Model Hubs and Frameworks (e.g., Hugging Face, PyTorch, TensorFlow) that provide access to pretrained models and development libraries.

• RAG and Orchestration Frameworks (e.g., LangChain, LlamaIndex, Haystack) that simplify building pipelines around LLMs.

• Vector Databases and Embedding Tools (e.g., FAISS, Milvus, Weaviate, Chroma) that enable semantic search and retrieval.

The Hugging Face Model Hub is a central repository that hosts thousands of pretrained AI models across domains such as natural language processing, computer vision, speech, and multimodal tasks. It enables researchers and developers to share, discover, and reuse models, reducing the time and computational cost required to train models from scratch.

The repository supports a wide range of architectures, including BERT, GPT, T5, CLIP, Stable Diffusion, and domain-specific models fine-tuned for tasks like sentiment analysis, machine translation, summarization, or image classification. Each model entry typically includes documentation, usage examples, licensing details, and community-driven updates.

A key strength of the Hub is its interoperability with Hugging Face libraries, allowing developers to easily load and deploy models with just a few lines of code. Additionally, the repository fosters an open research community, where contributors publish checkpoints, benchmarks, and evaluations, driving rapid innovation and transparency in AI development.

In this lecture, you will discover how the Hugging Face Datasets ecosystem supports different types of fine-tuning datasets and how its built-in Data Studio (Dataset Viewer + SQL console) helps you inspect, filter, and query dataset content—all before downloading or processing it.

You’ll learn:

• Common dataset formats and examples used in fine-tuning (e.g. instruction/response, classification, image/text pairs)

• How to explore dataset structure, metadata, and distributions via the Hugging Face Dataset Viewer

• Using filtering, searching, and histograms to understand class balance, missing values, ranges, and data types

• Running SQL queries directly inside the Dataset Viewer (via the SQL Console) to slice, transform, or extract subsets of data Hugging Face

• How large datasets are auto-converted to Parquet for efficient querying in Data Studio, and the constraints (e.g. 5GB view limits) Hugging Face+2Hugging Face+2

• Best practices when selecting or preparing datasets for fine-tuning based on what you observe in Data Studio