Description
Large Language Models (LLMs) have become the backbone of modern AI systems, powering applications from natural language understanding to complex, multi-agent systems. This Masterclass explores advanced techniques for designing, optimizing, and deploying LLM-based systems, with a particular focus on agentic architectures, retrieval augmented generation (RAG), advanced reasoning strategies, and hands-on implementations. Participants will gain practical knowledge of cutting-edge tools and frameworks, including LangGraph Studio, and will work with Jupyter notebooks provided for live coding demonstrations and experimentation.
While the hands-on sessions will focus on RAG, agentic setups, and deployment, attendees will receive notebooks covering all topics, including fine-tuning, optimization, and more, enabling them to continue exploring independently after the workshop.
Topics
Introduction to LLM Architectures
Overview of transformer models, including attention mechanisms and embeddings.
Key distinctions between Encoder-Only, Decoder-Only, and Encoder-Decodertransformers.
Optimization Techniques
Fine-tuning strategies, including PEFT, LoRA QLoRA.
Quantization and sharding to optimize performance and reduce resource consumption.
Sampling and decoding methods
Prompt Engineering
Retrieval-Augmented Generation (RAG)
Fundamentals of RAG: how it combines retrieval mechanisms with generative models.
Five levels of text splitting: effective strategies to improve performance of your languagemodel applications.
Advanced RAG architectures, including Corrective RAG, Self-RAG and Fusion RAG.
How to generate an evaluation dataset and improve and monitor the RAG system.
RAG pain points and how to solve them.
Introduction to Agents
Core principles of agentic LLMs and their applications.
One-step agent architectures and tool calling.
Introduction to LangGraph Studio for debugging and refining agent workflows.
Advanced Multi-Agent Architectures and Systems
Advanced Agent Architectures:
Language Agent Tree Search (LATS): Use reflection and reward-driven Monte Carlotree searches to explore agent actions.
Planning and Execution Agents: Implement basic planning agents capable of executing a series of tasks.
Advanced Reflection: Prompt agents to reflect on and revise outputs for improved reasoning.
Reflection: Guide agents to critique missing or superfluous details in their responses.
Self-Discovering Agents: Analyze and design agents capable of learning about and optimizing their own capabilities.
Building Multi-Agent Systems:
Specialized agents for retrieval, query transformation, hallucination checking, andhelpfulness assessment and complex multi-step tasks.
Visualizing data with code execution agents in secure sandbox environments likeE2B.
Deploying and Monitoring Agentic Systems
Hands-on implementation of agentic systems using:
Streamlit for creating interactive front-end interfaces.
DigitalOcean for secure deployment and real-time monitoring.
LangGraph Studio for debugging, visualization, and refinement of agent architectures.
Secure sandbox environments (e.g., E2B) for coding and execution agent architecture.
Using LangSmith and LangFuse to track agents and LLMs.
Practical Applications
Building intelligent systems for summarizing, retrieving, and generating context-sensitiveresponses.
Deploying agents for decision-making, task planning, and multi-step problem-solving.
Implementing LLMs for sentiment analysis, entity recognition, and more.
Responsible AI Practices
Strategies for maintaining AI transparency, explainability, and fairness.
Safeguards for ethical deployment of agent-based systems.
Requirements
Intermediate-level Python programming experience.
Familiarity with basic machine learning concepts, including neural networks and transformers helps, but is not a must, the basics of transformer will be covered.
Target Audience
This workshop is designed for AI practitioners, data scientists, and developers seeking to build advanced agent-based systems powered by LLMs. Participants will learn how to design, implement, and optimize intelligent agents for real-world applications, leveraging state-of-the-art tools and frameworks.
Speaker
Nicole Königstein
Chief AI Officer & Head of AI at quantmate, Author
