• Presentation of the final AI assistant
• Analysis of important technical aspects, options for improving response accuracy
• Analysis of architecture and how it affects the solution cost
• Discussion of next steps for improving the solution and implementation into business processes

• Cost of AI solutions: how to calculate expenses
• Best practices: when not to use LLM, RAG and agents
• Performance issues in interaction with LLM and agents
• Working with security and potentially dangerous actions of AI agents: guardrails, sandboxing, manual judgment, feedback loop, checkpoints
• Observability for AI application

Practice: 

Working with LLMOps systems, evaluating performance and cost of LLM usage

• Coding agents: hype or working tool
• AI assistants for the full development cycle: from working on an idea to deployment to production
• Cursor, Claude, Cline and other popular tools
• MCP servers, Skills and their integration into the development process

Practice: 

Creating an application entirely using a coding agent. Connecting MCP servers and Skills to extend agent capabilities.

• MCPs and tool calling
• Agent frameworks
• Agentic RAG
• Designing reliable AI agents
• Design patterns for agents
• Context engineering and memory management for agents

Practice: 

Creating a solution using agent framework. Planning and tool calling. Agentic RAG. Using user feedback.

• What is RAG and where it is used in the industry
• Why do we use RAG instead of training models?
• RAG architecture: retrieval algorithms and their optimization, response generation
• Context and memory: key things for building effective RAG solutions

Practice: 

Creating a chatbot for answering questions from your own (internal) data. Building RAG for large volumes of data when it's impossible to fit all information into the context window.

• Concepts and basic principles
• Similarity search, clustering, semantic and hybrid search, reranking
• Chunking strategies
• Embeddings for text, images, audio, video and multimodal embeddings
• Practical use cases: normalization and deduplication

Practice: 

Data vectorization. Similarity search with context consideration, normalization and deduplication.

• The probabilistic nature of LLM
• Writing scoring functions for AI solutions
• LLM-as-a-judge: using models to evaluate results
• Building a comprehensive evaluation pipeline as part of CI/CD

Practice:

Building a pipeline for evaluating AI solution performance. Preparing a test dataset and designing evaluation functions.

• The difference between zero-shot and few-shot prompting
• Optimizing prompts for a specific task
• CoT prompting and using models for reasoning
• How to write prompts to reduce hallucinations in LLM
• Security risks and ways to minimize them
• Versioning and testing prompts

Practice: 

Building a pipeline for extracting structured data from text files

• Training data, training, post-training and fine-tuning: what each stage is needed for
• Model architecture and its size
• Sampling and its impact on results
• Specialized models: coding, image processing, audio and video generation
• On-device and small (mini) models

Practice: 

Running and using local models for different use cases: coding, image generation, speech-to-text